Uploaded by Muhammad Akoodie

Introductory Econometrics 7E 2020

advertisement
Wooldridge
Introductory
Introductory
Econometrics
A Modern Approach
7e
Econometrics
A Modern Approach
7e
Jeffrey M. Wooldridge
SE/Author/Author, Title Xe ISBN-13: 978XXXXXXXXXX ©2015 Designer:
Printer: Binding: Casebound Trim: 8" x 10" CMYK
Introductory
Econometrics
A Modern Approach
Seventh Edition
Jeffrey M. Wooldridge
Michigan State University
Australia • Brazil • Mexico • Singapore • United Kingdom • United States
58860_fm_hr_i-xxii.indd 1
10/23/18 6:11 PM
Introductory Econometrics: A Modern
Approach, Seventh Edition
Jeffrey M. Wooldridge
© 2020, 2016 Cengage Learning, Inc.
Unless otherwise noted, all content is © Cengage.
ALL RIGHTS RESERVED. No part of this work covered by the
copyright herein may be reproduced or distributed in any form or
Senior Vice President, Higher Education Product
Management: Erin Joyner
by any means, except as permitted by U.S. copyright law, without
the prior written permission of the copyright owner.
Product Director: Jason Fremder
Sr. Product Manager: Michael Parthenakis
For product information and technology assistance, contact us at
Cengage Customer & Sales Support, 1-800-354-9706
Sr. Learning Designer: Sarah Keeling
or support.cengage.com.
Sr. Content Manager: Anita Verma
In-House Subject Matter Expert (s): Eugenia
For permission to use material from this text or product, submit all
requests online at www.cengage.com/permissions.
Belova, Ethan Crist and Kasie Jean
Digital Delivery Lead: Timothy Christy
Product Assistant: Matt Schiesl
Manufacturing Planner: Kevin Kluck
Library of Congress Control Number: 2018956380
ISBN: 978-1-337-55886-0
Production Service: SPi-Global
Intellectual Property
Analyst: Jennifer Bowes
Project Manager: Julie Geagan-Chevez
Marketing Manager: John Carey
Sr. Designer: Bethany Bourgeois
Cover Designer: Tin Box Studio
Cengage
20 Channel Center Street
Boston, MA 02210
USA
Cengage is a leading provider of customized learning solutions
with employees residing in nearly 40 different countries and sales
in more than 125 countries around the world. Find your local
representative at www.cengage.com.
Cengage products are represented in Canada by
Nelson Education, Ltd.
To learn more about Cengage platforms and services, register or
access your online learning solution, or purchase materials for
your course, visit www.cengage.com.
Printed in the United States of America
Print Number: 01
Print Year: 2018
58860_fm_hr_i-xxii.indd 2
10/23/18 6:11 PM
Brief Contents
Chapter 1
The Nature of Econometrics and Economic Data
Part 1: Regression Analysis with Cross-Sectional Data
Chapter 2
Chapter 3
Chapter 4
Chapter 5
Chapter 6
Chapter 7
Chapter 8
Chapter 9
1
19
The Simple Regression Model
20
Multiple Regression Analysis: Estimation
66
Multiple Regression Analysis: Inference
117
Multiple Regression Analysis: OLS Asymptotics
163
Multiple Regression Analysis: Further Issues
181
Multiple Regression Analysis with Qualitative Information
220
Heteroskedasticity262
More on Specification and Data Issues
294
Part 2: Regression Analysis with Time Series Data
333
Chapter 10
Chapter 11
Chapter 12
334
366
394
Basic Regression Analysis with Time Series Data
Further Issues in Using OLS with Time Series Data
Serial Correlation and Heteroskedasticity in Time Series Regressions
Part 3: Advanced Topics
425
Chapter 13
Chapter 14
Chapter 15
Chapter 16
Chapter 17
Chapter 18
Chapter 19
426
462
495
534
559
604
642
Pooling Cross Sections across Time: Simple Panel Data Methods
Advanced Panel Data Methods
Instrumental Variables Estimation and Two-Stage Least Squares
Simultaneous Equations Models
Limited Dependent Variable Models and Sample Selection Corrections
Advanced Time Series Topics
Carrying Out an Empirical Project
Appendices
Math Refresher A Basic Mathematical Tools
Math Refresher B Fundamentals of Probability
Math Refresher C Fundamentals of Mathematical Statistics
Advanced Treatment D Summary of Matrix Algebra
Advanced Treatment E The Linear Regression Model in Matrix Form
666
684
714
749
760
Answers to Going Further Questions
775
Statistical Tables
784
References791
Glossary797
Index812
iii
58860_fm_hr_i-xxii.indd 3
10/23/18 6:11 PM
Contents
Preface
xii
About the Author
2-4 Units of Measurement and Functional Form 36
2-4a The Effects of Changing Units of Measurement
on OLS Statistics 36
2-4b Incorporating Nonlinearities in Simple
Regression 37
2-4c The Meaning of “Linear” Regression 40
xxii
The Nature of Econometrics
and Economic Data 1
chapter 1
1-1 What Is Econometrics?
2-5 Expected Values and Variances of the OLS
Estimators 40
2-5a Unbiasedness of OLS 40
2-5b Variances of the OLS Estimators 45
2-5c Estimating the Error Variance 48
1
1-2 Steps in Empirical Economic Analysis
2
1-3 The Structure of Economic Data 5
1-3a Cross-Sectional Data 5
1-3b Time Series Data 7
1-3c Pooled Cross Sections 8
1-3d Panel or Longitudinal Data 9
1-3e A Comment on Data Structures 10
2-6 Regression through the Origin and Regression
on a Constant 50
1-4 Causality, Ceteris Paribus, and Counterfactual
Reasoning 10
2-7 Regression on a Binary Explanatory
Variable 51
2-7a Counterfactual Outcomes, Causality, and Policy
Analysis 53
Summary
Summary
14
Key Terms
Problems
15
15
Computer Exercises
Problems
15
57
chapter 3
Estimation
Regression Analysis with
­Cross-Sectional Data 19
The Simple Regression Model
2-1 Definition of the Simple Regression Model
58
Computer Exercises
Part 1
chapter 2
56
Key Terms
20
20
2-2 Deriving the Ordinary Least Squares Estimates
2-2a A Note on Terminology 31
2-3 Properties of OLS on Any Sample of Data 32
2-3a Fitted Values and Residuals 32
2-3b Algebraic Properties of OLS Statistics 32
2-3c Goodness-of-Fit 35
24
62
Multiple Regression Analysis:
66
3-1 Motivation for Multiple Regression 67
3-1a The Model with Two Independent Variables 67
3-1b The Model with k Independent Variables 69
3-2 Mechanics and Interpretation of Ordinary Least
Squares 70
3-2a Obtaining the OLS Estimates 70
3-2b Interpreting the OLS Regression Equation 71
3-2c On the Meaning of “Holding Other Factors Fixed”
in Multiple Regression 73
3-2d Changing More Than One Independent Variable
Simultaneously 74
iv
58860_fm_hr_i-xxii.indd 4
10/23/18 6:11 PM
v
Contents
3-2e OLS Fitted Values and Residuals 74
3-2f A “Partialling Out” Interpretation of Multiple
Regression 75
3-2g Comparison of Simple and Multiple Regression
Estimates 75
3-2h Goodness-of-Fit 76
3-2i Regression through the Origin 79
3-3 The Expected Value of the OLS Estimators 79
3-3a Including Irrelevant Variables in a Regression
Model 83
3-3b Omitted Variable Bias: The Simple Case 84
3-3c Omitted Variable Bias: More General Cases 87
3-4 The Variance of the OLS Estimators 87
3-4a The Components of the OLS Variances:
Multicollinearity 89
3-4b Variances in Misspecified Models 92
3-4c Estimating s2: Standard Errors of the OLS
Estimators 93
3-5 Efficiency of OLS: The Gauss-Markov
Theorem 95
3-6 Some Comments on the Language of Multiple
Regression Analysis 96
3-7 Several Scenarios for Applying Multiple
Regression 97
3-7a Prediction 98
3-7b Efficient Markets 98
3-7c Measuring the Tradeoff between Two
Variables 99
3-7d Testing for Ceteris Paribus Group
Differences 99
3-7e Potential Outcomes, Treatment Effects, and Policy
Analysis 100
Summary
102
Key Terms
Problems
104
4-3 Confidence Intervals 134
4-4 Testing Hypotheses about a Single Linear
Combination of the Parameters 136
4-5 Testing Multiple Linear Restrictions: The
F Test 139
4-5a Testing Exclusion Restrictions 139
4-5b Relationship between F and t Statistics 144
4-5c The R-Squared Form of the F Statistic 145
4-5d Computing p-values for F Tests 146
4-5e The F Statistic for Overall Significance of a
Regression 147
4-5f Testing General Linear Restrictions 148
4-6 Reporting Regression Results 149
4-7 Revisiting Causal Effects and Policy
Analysis 151
Summary
152
Key Terms
Problems
154
154
Computer Exercises
Multiple Regression Analysis: OLS
Asymptotics 163
chapter 5
5-1 Consistency 164
5-1a Deriving the Inconsistency in OLS
109
Multiple Regression Analysis:
Inference 117
chapter 4
4-1 Sampling Distributions of the OLS
Estimators 117
4-2 Testing Hypotheses about a Single Population
Parameter: The t Test 120
4-2a Testing against One-Sided Alternatives 122
4-2b Two-Sided Alternatives 126
4-2c Testing Other Hypotheses about bj 128
4-2d Computing p-Values for t Tests 130
167
5-2 Asymptotic Normality and Large Sample
Inference 168
5-2a Other Large Sample Tests: The Lagrange
Multiplier Statistic 172
Summary
Problems
175
176
Key Terms
58860_fm_hr_i-xxii.indd 5
159
5-3 Asymptotic Efficiency of OLS
104
Computer Exercises
4-2e A Reminder on the Language of Classical
Hypothesis Testing 132
4-2f Economic, or Practical, versus Statistical
Significance 132
176
176
Computer Exercises
178
Multiple Regression Analysis:
Further Issues 181
chapter 6
6-1 Effects of Data Scaling on OLS Statistics 181
6-1a Beta Coefficients 184
6-2 More on Functional Form 186
6-2a More on Using Logarithmic Functional
Forms 186
10/23/18 6:11 PM
vi
Contents
6-2b Models with Quadratics 188
6-2c Models with Interaction Terms 192
6-2d Computing Average Partial Effects 194
6-3 More on Goodness-of-Fit and Selection of
Regressors 195
6-3a Adjusted R-Squared 196
6-3b Using Adjusted R-Squared to Choose between
Nonnested Models 197
6-3c Controlling for Too Many Factors in Regression
Analysis 199
6-3d Adding Regressors to Reduce the Error
Variance 200
6-4 Prediction and Residual Analysis 201
6.4a Confidence Intervals for Predictions 201
6-4b Residual Analysis 205
6-4c Predicting y When log(y) Is the Dependent
Variable 205
6-4d Predicting y When the Dependent Variable Is
log(y) 207
Summary
209
Key Terms
Problems
211
211
Computer Exercises
214
Multiple Regression Analysis with
Qualitative Information 220
chapter 7
7-7 Interpreting Regression Results with Discrete
Dependent Variables 249
Summary
Key Terms
Problems
221
7-2 A Single Dummy Independent
Variable 222
7-2a Interpreting Coefficients on Dummy
Explanatory Variables When the Dependent
Variable Is log(y) 226
7-3 Using Dummy Variables for Multiple
Categories 228
7-3a Incorporating Ordinal Information by Using
Dummy Variables 230
7-4 Interactions Involving Dummy Variables 232
7-4a Interactions among Dummy Variables 232
7-4b Allowing for Different Slopes 233
7-4c Testing for Differences in Regression Functions
across Groups 237
7-5 A Binary Dependent Variable: The Linear
Probability Model 239
7-6 More on Policy Analysis and Program
Evaluation 244
7-6a Program Evaluation and Unrestricted Regression
Adjustment 245
58860_fm_hr_i-xxii.indd 6
251
251
Computer Exercises
chapter 8
256
Heteroskedasticity
262
8-1 Consequences of Heteroskedasticity for OLS
262
8-2 Heteroskedasticity-Robust Inference after OLS
Estimation 263
8-2a Computing Heteroskedasticity-Robust LM
Tests 267
8-3 Testing for Heteroskedasticity 269
8-3a The White Test for Heteroskedasticity
271
8-4 Weighted Least Squares Estimation 273
8-4a The Heteroskedasticity Is Known up to a
Multiplicative Constant 273
8-4b The Heteroskedasticity Function
Must Be Estimated: Feasible GLS 278
8-4c What If the Assumed Heteroskedasticity Function Is
Wrong? 281
8-4d Prediction and Prediction Intervals with
Heteroskedasticity 283
8-5 The Linear Probability Model Revisited 284
Summary
7-1 Describing Qualitative Information
250
286
Key Terms
Problems
287
287
Computer Exercises
290
More on Specification and
Data Issues 294
chapter 9
9-1 Functional Form Misspecification 295
9-1a RESET as a General Test for Functional
Form Misspecification 297
9-1b Tests against Nonnested Alternatives 298
9-2 Using Proxy Variables for Unobserved Explanatory
Variables 299
9-2a Using Lagged Dependent Variables as Proxy
Variables 303
9-2b A Different Slant on Multiple Regression 304
9-2c Potential Outcomes and Proxy Variables 305
9-3 Models with Random Slopes
306
9-4 Properties of OLS under Measurement Error 308
9-4a Measurement Error in the Dependent Variable 308
10/23/18 6:11 PM
vii
Contents
Problems
9-4b Measurement Error in an Explanatory
Variable 310
9-5 Missing Data, Nonrandom Samples, and Outlying
Observations 313
9-5a Missing Data 313
9-5b Nonrandom Samples 315
9-5c Outliers and Influential Observations 317
9-6 Least Absolute Deviations Estimation
Summary
Key Terms
Problems
321
323
324
324
Computer Exercises
328
Regression Analysis with Time
Series Data 333
Basic Regression Analysis with
Time Series Data 334
chapter 10
10-1 The Nature of Time Series Data
334
10-2 Examples of Time Series Regression
Models 335
10-2a Static Models 336
10-2b Finite Distributed Lag Models 336
10-2c A Convention about the Time Index 338
10-3 Finite Sample Properties of OLS under Classical
Assumptions 339
10-3a Unbiasedness of OLS 339
10-3b The Variances of the OLS Estimators and the
Gauss-Markov Theorem 342
10-3c Inference under the Classical Linear Model
Assumptions 344
10-4 Functional Form, Dummy Variables, and Index
Numbers 345
10-5 Trends and Seasonality 351
10-5a Characterizing Trending Time Series 351
10-5b Using Trending Variables in Regression
Analysis 354
10-5c A Detrending Interpretation of Regressions
with a Time Trend 356
10-5d Computing R-Squared When the Dependent
Variable Is Trending 357
10-5e Seasonality 358
Key Terms
58860_fm_hr_i-xxii.indd 7
360
361
363
Further Issues in Using OLS with
Time Series Data 366
chapter 11
11-1 Stationary and Weakly Dependent Time
Series 367
11-1a Stationary and Nonstationary Time Series
11-1b Weakly Dependent Time Series 368
11-2 Asymptotic Properties of OLS
Part 2
Summary
361
Computer Exercises
367
370
11-3 Using Highly Persistent Time Series in Regression
Analysis 376
11-3a Highly Persistent Time Series 376
11-3b Transformations on Highly Persistent Time
Series 380
11-3c Deciding Whether a Time Series Is I(1) 381
11-4 Dynamically Complete Models and the Absence of
Serial Correlation 382
11-5 The Homoskedasticity Assumption for Time
Series Models 385
Summary
Key Terms
Problems
386
387
387
Computer Exercises
390
Serial Correlation and
Heteroskedasticity in Time Series
Regressions 394
chapter 12
12-1 Properties of OLS with Serially Correlated
Errors 395
12-1a Unbiasedness and Consistency 395
12-1b Efficiency and Inference 395
12-1c Goodness-of-Fit 396
12-1d Serial Correlation in the Presence
of Lagged Dependent Variables 396
12-2 Serial Correlation–Robust Inference
after OLS 398
12-3 Testing for Serial Correlation 401
12-3a A t Test for AR(1) Serial Correlation with
Strictly Exogenous Regressors 402
12-3b The Durbin-Watson Test under Classical
Assumptions 403
12-3c Testing for AR(1) Serial Correlation without
Strictly Exogenous Regressors 404
12-3d Testing for Higher-Order Serial Correlation 406
10/23/18 6:11 PM
viii
Contents
12-4 Correcting for Serial Correlation with Strictly
Exogenous Regressors 407
12-4a Obtaining the Best Linear Unbiased
Estimator in the AR(1) Model 408
12-4b Feasible GLS Estimation with AR(1)
Errors 409
12-4c Comparing OLS and FGLS 411
12-4d Correcting for Higher-Order Serial
Correlation 413
12-4e What if the Serial Correlation Model Is
Wrong? 413
12-5 Differencing and Serial Correlation
414
12-6 Heteroskedasticity in Time Series
Regressions 415
12-6a Heteroskedasticity-Robust Statistics 416
12-6b Testing for Heteroskedasticity 416
12-6c Autoregressive Conditional
Heteroskedasticity 417
12-6d Heteroskedasticity and Serial Correlation in
Regression Models 418
Summary
Key Terms
Problems
419
420
Key Terms
Problems
452
452
Computer Exercises
453
Advanced Panel
Data Methods 462
chapter 14
14-1 Fixed Effects Estimation 463
14-1a The Dummy Variable Regression 466
14-1b Fixed Effects or First Differencing? 467
14-1c Fixed Effects with Unbalanced Panels 468
14-2 Random Effects Models 469
14-2a Random Effects or Pooled OLS? 473
14-2b Random Effects or Fixed Effects? 473
14-3 The Correlated Random Effects Approach
14-3a Unbalanced Panels 476
474
14-4 General Policy Analysis with Panel Data 477
14-4a Advanced Considerations with Policy
Analysis 478
Summary
421
Key Terms
Problems
Part 3
Advanced Topics
425
Pooling Cross Sections across
Time: Simple Panel Data Methods 426
chapter 13
13-1 Pooling Independent Cross Sections across
Time 427
13-1a The Chow Test for Structural Change across
Time 431
13-2 Policy Analysis with Pooled Cross Sections 431
13-2a Adding an Additional Control Group 436
13-2b A General Framework for Policy Analysis
with Pooled Cross Sections 437
13-3 Two-Period Panel Data Analysis 439
13-3a Organizing Panel Data 444
13-4 Policy Analysis with Two-Period Panel
Data 444
13-5 Differencing with More Than Two Time
Periods 447
13-5a Potential Pitfalls in First Differencing Panel
Data 451
58860_fm_hr_i-xxii.indd 8
451
14-5 Applying Panel Data Methods to Other Data
Structures 480
420
Computer Exercises
Summary
483
484
484
Computer Exercises
486
Instrumental Variables Estimation
and Two-Stage Least Squares 495
chapter 15
15-1 Motivation: Omitted Variables in a Simple
Regression Model 496
15-1a Statistical Inference with the IV
Estimator 500
15-1b Properties of IV with a Poor Instrumental
Variable 503
15-1c Computing R-Squared after IV Estimation
505
15-2 IV Estimation of the Multiple Regression
Model 505
15-3 Two-Stage Least Squares 509
15-3a A Single Endogenous Explanatory
Variable 509
15-3b Multicollinearity and 2SLS 511
15-3c Detecting Weak Instruments 512
15-3d Multiple Endogenous Explanatory
Variables 513
15-3e Testing Multiple Hypotheses after 2SLS
Estimation 513
10/23/18 6:11 PM
ix
Contents
15-4 IV Solutions to Errors-in-Variables Problems
514
15-5 Testing for Endogeneity and Testing Overidentifying
Restrictions 515
15-5a Testing for Endogeneity 515
15-5b Testing Overidentification Restrictions 516
15-6 2SLS with Heteroskedasticity
518
15-7 Applying 2SLS to Time Series Equations 519
15-8 Applying 2SLS to Pooled Cross Sections
and Panel Data 521
Summary
Key Terms
Problems
522
523
523
Computer Exercises
17-2 The Tobit Model for Corner Solution
Responses 571
17-2a Interpreting the Tobit Estimates 572
17-2b Specification Issues in Tobit Models 578
17-3 The Poisson Regression Model 578
17-4 Censored and Truncated Regression Models 582
17-4a Censored Regression Models 583
17-4b Truncated Regression Models 586
17-5 Sample Selection Corrections 588
17-5a When Is OLS on the Selected Sample
Consistent? 588
17-5b Incidental Truncation 589
Summary
526
Key Terms
Problems
593
593
594
Simultaneous Equations
Models 534
Computer Exercises
16-1 The Nature of Simultaneous Equations
Models 535
chapter 18
16-2 Simultaneity Bias in OLS
18-1 Infinite Distributed Lag Models 605
18-1a The Geometric (or Koyck) Distributed Lag
Model 607
18-1b Rational Distributed Lag Models 608
chapter 16
538
16-3 Identifying and Estimating a Structural
Equation 539
16-3a Identification in a Two-Equation System 540
16-3b Estimation by 2SLS 543
16-4 Systems with More Than Two Equations 545
16-4a Identification in Systems with Three or More
Equations 545
16-4b Estimation 546
16-5 Simultaneous Equations Models with Time
Series 546
16-6 Simultaneous Equations Models with Panel
Data 549
Summary
Key Terms
Problems
551
552
552
Computer Exercises
555
Limited Dependent Variable Models
and Sample Selection Corrections 559
chapter 17
Advanced Time Series Topics
58860_fm_hr_i-xxii.indd 9
604
18-2 Testing for Unit Roots 610
18-3 Spurious Regression 614
18-4 Cointegration and Error Correction Models 616
18-4a Cointegration 616
18-4b Error Correction Models 620
18-5 Forecasting 622
18-5a Types of Regression Models Used for
Forecasting 623
18-5b One-Step-Ahead Forecasting 624
18-5c Comparing One-Step-Ahead Forecasts 627
18-5d Multiple-Step-Ahead Forecasts 628
18-5e Forecasting Trending, Seasonal, and Integrated
Processes 631
Summary
Key Terms
Problems
635
636
636
Computer Exercises
17-1 Logit and Probit Models for Binary
Response 560
17-1a Specifying Logit and Probit Models 560
17-1b Maximum Likelihood Estimation of Logit and
Probit Models 563
17-1c Testing Multiple Hypotheses 564
17-1d Interpreting the Logit and Probit Estimates 565
596
638
Carrying Out an Empirical
Project 642
chapter 19
19-1 Posing a Question
642
19-2 Literature Review 644
10/23/18 6:11 PM
x
Contents
19-3 Data Collection 645
19-3a Deciding on the Appropriate Data Set 645
19-3b Entering and Storing Your Data 646
19-3c Inspecting, Cleaning, and Summarizing Your
Data 647
19-4 Econometric Analysis
648
19-5 Writing an Empirical Paper 651
19-5a Introduction 651
19-5b Conceptual (or Theoretical)
Framework 652
19-5c Econometric Models and Estimation
Methods 652
19-5d The Data 654
19-5e Results 655
19.5f Conclusions 656
19-5g Style Hints 656
Summary
658
658
Key Terms
Sample Empirical Projects
List of Journals
Data Sources
658
664
665
Math Refresher A Basic Mathematical
Tools 666
A-1 The Summation Operator and Descriptive
Statistics 666
A-2 Properties of Linear Functions
A-3 Proportions and Percentages
668
671
A-4 Some Special Functions and Their
Properties 672
A-4a Quadratic Functions 672
A-4b The Natural Logarithm 674
A-4c The Exponential Function 677
A-5 Differential Calculus
Summary
B-3 Features of Probability Distributions 691
B-3a A Measure of Central Tendency: The Expected
Value 691
B-3b Properties of Expected Values 692
B-3c Another Measure of Central Tendency: The
Median 694
B-3d Measures of Variability: Variance and Standard
Deviation 695
B-3e Variance 695
B-3f Standard Deviation 696
B-3g Standardizing a Random Variable 696
B-3h Skewness and Kurtosis 697
B-4 Features of Joint and Conditional
Distributions 697
B-4a Measures of Association: Covariance and
Correlation 697
B-4b Covariance 697
B-4c Correlation Coefficient 698
B-4d Variance of Sums of Random Variables 699
B-4e Conditional Expectation 700
B-4f Properties of Conditional Expectation 702
B-4g Conditional Variance 704
B-5 The Normal and Related Distributions 704
B-5a The Normal Distribution 704
B-5b The Standard Normal Distribution 705
B-5c Additional Properties of the Normal
Distribution 707
B-5d The Chi-Square Distribution 708
B-5e The t Distribution 708
B-5f The F Distribution 709
Summary
Key Terms
Problems
680
Key Terms
Problems
678
B-2 Joint Distributions, Conditional Distributions,
and Independence 688
B-2a Joint Distributions and Independence 688
B-2b Conditional Distributions 690
711
711
711
681
Math Refresher C Fundamentals of
681
Math Refresher B
Probability
684
Fundamentals of
B-1 Random Variables and Their Probability
Distributions 684
B-1a Discrete Random Variables 685
B-1b Continuous Random Variables 687
58860_fm_hr_i-xxii.indd 10
Mathematical Statistics
714
C-1 Populations, Parameters, and Random
Sampling 714
C-1a Sampling 714
C-2 Finite Sample Properties of Estimators
C-2a Estimators and Estimates 715
C-2b Unbiasedness 716
715
10/23/18 6:11 PM
Contents
C-2c The Sampling Variance of Estimators 718
C-2d Efficiency 719
D-2e Partitioned Matrix Multiplication 752
D-2f Trace 753
D-2g Inverse 753
C-3 Asymptotic or Large Sample Properties of
Estimators 721
C-3a Consistency 721
C-3b Asymptotic Normality 723
C-4 General Approaches to Parameter Estimation
C-4a Method of Moments 725
C-4b Maximum Likelihood 725
C-4c Least Squares 726
D-3 Linear Independence and Rank of a
Matrix 754
724
D-4 Quadratic Forms and Positive Definite
Matrices 754
D-5 Idempotent Matrices 755
D-6 Differentiation of Linear and Quadratic
Forms 755
C-5 Interval Estimation and Confidence Intervals 727
C-5a The Nature of Interval Estimation 727
C-5b Confidence Intervals for the Mean from a Normally
Distributed Population 729
C-5c A Simple Rule of Thumb for a 95% Confidence
Interval 731
C-5d Asymptotic Confidence Intervals for Nonnormal
Populations 732
D-7 Moments and Distributions of Random
Vectors 756
D-7a Expected Value 756
D-7b Variance-Covariance Matrix 756
D-7c Multivariate Normal Distribution 756
D-7d Chi-Square Distribution 757
D-7e t Distribution 757
D-7f F Distribution 757
C-6 Hypothesis Testing 733
C-6a Fundamentals of Hypothesis Testing 733
C-6b Testing Hypotheses about the Mean in a Normal
Population 735
C-6c Asymptotic Tests for Nonnormal
Populations 738
C-6d Computing and Using p-Values 738
C-6e The Relationship between Confidence Intervals
and Hypothesis Testing 741
C-6f Practical versus Statistical Significance 742
Summary
C-7 Remarks on Notation
Summary
Key Terms
Problems
744
749
D-1 Basic Definitions
Summary of Matrix
757
758
Problems
Advanced Treatment E The Linear Regression
Model in Matrix Form
760
E-1 The Model and Ordinary Least
Squares Estimation 760
E-1a The Frisch-Waugh Theorem 762
E-2 Finite Sample Properties of OLS
763
E-4 Some Asymptotic Analysis 769
E-4a Wald Statistics for Testing Multiple
Hypotheses 771
Summary
771
Key Terms
749
D-2 Matrix Operations 750
D-2a Matrix Addition 750
D-2b Scalar Multiplication 750
D-2c Matrix Multiplication 751
D-2d Transpose 752
58860_fm_hr_i-xxii.indd 11
757
Key Terms
E-3 Statistical Inference 767
744
Advanced Treatment D Algebra
743
743
xi
771
Problems
772
Answers to Going Further Questions
Statistical Tables
References
Glossary
Index
775
784
791
797
812
10/23/18 6:11 PM
Preface
In ALL content, please indent the first paragraph as well, like the following ones. My motivation
for writing the first edition of Introductory Econometrics: A Modern Approach was that I saw a fairly
wide gap between how econometrics is taught to undergraduates and how empirical researchers think
about and apply econometric methods. I became convinced that teaching introductory econometrics
from the perspective of professional users of econometrics would actually simplify the presentation,
in addition to making the subject much more interesting.
Based on the positive reactions to the several earlier editions, it appears that my hunch was correct.
Many instructors, having a variety of backgrounds and interests and teaching students with different
levels of preparation, have embraced the modern approach to econometrics espoused in this text. The
emphasis in this edition is still on applying econometrics to real-world problems. Each econometric
method is motivated by a particular issue facing researchers analyzing nonexperimental data. The focus
in the main text is on understanding and interpreting the assumptions in light of actual empirical applications: the mathematics required is no more than college algebra and basic probability and statistics.
Designed for Today’s Econometrics Course
The seventh edition preserves the overall organization of the sixth. The most noticeable feature
that distinguishes this text from most others is the separation of topics by the kind of data being analyzed. This is a clear departure from the traditional approach, which presents a linear model, lists all
assumptions that may be needed at some future point in the analysis, and then proves or asserts results
without clearly connecting them to the assumptions. My approach is first to treat, in Part 1, multiple regression analysis with cross-sectional data, under the assumption of random sampling. This
setting is natural to students because they are familiar with random sampling from a population in
their introductory statistics courses. Importantly, it allows us to distinguish assumptions made about
the underlying population regression model—assumptions that can be given economic or behavioral
­content—from assumptions about how the data were sampled. Discussions about the consequences of
nonrandom sampling can be treated in an intuitive fashion after the students have a good grasp of the
multiple regression model estimated using random samples.
An important feature of a modern approach is that the explanatory variables—along with the
dependent variable—are treated as outcomes of random variables. For the social sciences, allowing random explanatory variables is much more realistic than the traditional assumption of nonrandom explanatory variables. As a nontrivial benefit, the population model/random sampling approach
reduces the number of assumptions that students must absorb and understand. Ironically, the classical
approach to regression analysis, which treats the explanatory variables as fixed in repeated samples
and is still pervasive in introductory texts, literally applies to data collected in an experimental setting.
In addition, the contortions required to state and explain assumptions can be confusing to students.
My focus on the population model emphasizes that the fundamental assumptions underlying
regression analysis, such as the zero mean assumption on the unobservable error term, are properly
xii
58860_fm_hr_i-xxii.indd 12
10/23/18 6:11 PM
Preface
xiii
stated conditional on the explanatory variables. This leads to a clear understanding of the kinds of
problems, such as heteroskedasticity (nonconstant variance), that can invalidate standard inference
procedures. By focusing on the population, I am also able to dispel several misconceptions that arise
in econometrics texts at all levels. For example, I explain why the usual R-squared is still valid as a
goodness-of-fit measure in the presence of heteroskedasticity (Chapter 8) or serially correlated errors
(Chapter 12); I provide a simple demonstration that tests for functional form should not be viewed
as general tests of omitted variables (Chapter 9); and I explain why one should always include in a
regression model extra control variables that are uncorrelated with the explanatory variable of interest, which is often a key policy variable (Chapter 6).
Because the assumptions for cross-sectional analysis are relatively straightforward yet realistic, students can get involved early with serious cross-sectional applications without having to worry
about the thorny issues of trends, seasonality, serial correlation, high persistence, and spurious regression that are ubiquitous in time series regression models. Initially, I figured that my treatment of
regression with cross-sectional data followed by regression with time series data would find favor
with instructors whose own research interests are in applied microeconomics, and that appears to be
the case. It has been gratifying that adopters of the text with an applied time series bent have been
equally enthusiastic about the structure of the text. By postponing the econometric analysis of time
series data, I am able to put proper focus on the potential pitfalls in analyzing time series data that do
not arise with cross-sectional data. In effect, time series econometrics finally gets the serious treatment it deserves in an introductory text.
As in the earlier editions, I have consciously chosen topics that are important for reading journal
articles and for conducting basic empirical research. Within each topic, I have deliberately omitted
many tests and estimation procedures that, while traditionally included in textbooks, have not withstood the empirical test of time. Likewise, I have emphasized more recent topics that have clearly
demonstrated their usefulness, such as obtaining test statistics that are robust to heteroskedasticity
(or serial correlation) of unknown form, using multiple years of data for policy analysis, or solving
the omitted variable problem by instrumental variables methods. I appear to have made fairly good
choices, as I have received only a handful of suggestions for adding or deleting material.
I take a systematic approach throughout the text, by which I mean that each topic is presented by
building on the previous material in a logical fashion, and assumptions are introduced only as they
are needed to obtain a conclusion. For example, empirical researchers who use econometrics in their
research understand that not all of the Gauss-Markov assumptions are needed to show that the ordinary least squares (OLS) estimators are unbiased. Yet the vast majority of econometrics texts introduce a complete set of assumptions (many of which are redundant or in some cases even logically
conflicting) before proving the unbiasedness of OLS. Similarly, the normality assumption is often
included among the assumptions that are needed for the Gauss-Markov Theorem, even though it is
fairly well known that normality plays no role in showing that the OLS estimators are the best linear
unbiased estimators.
My systematic approach is illustrated by the order of assumptions that I use for multiple regression in Part 1. This structure results in a natural progression for briefly summarizing the role of each
assumption:
MLR.1: Introduce the population model and interpret the population parameters (which we hope
to estimate).
MLR.2: Introduce random sampling from the population and describe the data that we use to
estimate the population parameters.
MLR.3: Add the assumption on the explanatory variables that allows us to compute the estimates
from our sample; this is the so-called no perfect collinearity assumption.
MLR.4: Assume that, in the population, the mean of the unobservable error does not depend on the
values of the explanatory variables; this is the “mean independence” assumption combined with a
zero population mean for the error, and it is the key assumption that delivers unbiasedness of OLS.
58860_fm_hr_i-xxii.indd 13
10/23/18 6:11 PM
xiv
Preface
After introducing Assumptions MLR.1 to MLR.3, one can discuss the algebraic properties of ordinary least squares—that is, the properties of OLS for a particular set of data. By adding Assumption
MLR.4, we can show that OLS is unbiased (and consistent). Assumption MLR.5 (homoskedasticity) is added for the Gauss-Markov Theorem and for the usual OLS variance formulas to be valid.
Assumption MLR.6 (normality), which is not introduced until Chapter 4, is added to round out the
classical linear model assumptions. The six assumptions are used to obtain exact statistical inference
and to conclude that the OLS estimators have the smallest variances among all unbiased estimators.
I use parallel approaches when I turn to the study of large-sample properties and when I treat
regression for time series data in Part 2. The careful presentation and discussion of assumptions
makes it relatively easy to transition to Part 3, which covers advanced topics that include using pooled
cross-sectional data, exploiting panel data structures, and applying instrumental variables methods.
Generally, I have strived to provide a unified view of econometrics, where all estimators and test statistics are obtained using just a few intuitively reasonable principles of estimation and testing (which,
of course, also have rigorous justification). For example, regression-based tests for heteroskedasticity
and serial correlation are easy for students to grasp because they already have a solid understanding
of regression. This is in contrast to treatments that give a set of disjointed recipes for outdated econometric testing procedures.
Throughout the text, I emphasize ceteris paribus relationships, which is why, after one chapter on
the simple regression model, I move to multiple regression analysis. The multiple regression setting
motivates students to think about serious applications early. I also give prominence to policy analysis
with all kinds of data structures. Practical topics, such as using proxy variables to obtain ceteris paribus effects and interpreting partial effects in models with interaction terms, are covered in a simple
fashion.
Designed at Undergraduates, Applicable
to Master’s Students
The text is designed for undergraduate economics majors who have taken college algebra and
one-semester of introductory probability and statistics. (Math Refresher A, B, and C contain the
requisite background material.) A one-semester or one-quarter econometrics course would not be
expected to cover all, or even any, of the more advanced material in Part 3. A typical introductory course includes Chapters 1 through 8, which cover the basics of simple and multiple regression for cross-sectional data. Provided the emphasis is on intuition and interpreting the empirical
examples, the material from the first eight chapters should be accessible to undergraduates in most
economics departments. Most instructors will also want to cover at least parts of the c­ hapters
on regression analysis with time series data, Chapters 10 and 12, in varying degrees of depth.
In the one-semester course that I teach at Michigan State, I cover Chapter 10 fairly carefully,
give an overview of the material in Chapter 11, and cover the material on serial correlation in
Chapter 12. I find that this basic one-semester course puts students on a solid footing to write
empirical papers, such as a term paper, a senior seminar paper, or a senior thesis. Chapter 9
­c ontains more specialized topics that arise in analyzing cross-sectional data, including data
­problems such as outliers and nonrandom sampling; for a one-semester course, it can be skipped
without loss of continuity.
The structure of the text makes it ideal for a course with a cross-sectional or policy analysis
focus: the time series chapters can be skipped in lieu of topics from Chapters 9 or 15. The new material on potential outcomes added to the first nine chapters should help the instructor craft a course
that provides an introduction to modern policy analysis. Chapter 13 is advanced only in the sense
that it treats two new data structures: independently pooled cross sections and two-period panel data
analysis. Such data structures are especially useful for policy analysis, and the chapter provides
58860_fm_hr_i-xxii.indd 14
10/23/18 6:11 PM
Preface
xv
several examples. Students with a good grasp of Chapters 1 through 8 will have little difficulty with
Chapter 13. Chapter 14 covers more advanced panel data methods and would probably be covered
only in a second course. A good way to end a course on cross-sectional methods is to cover the rudiments of instrumental variables estimation in Chapter 15.
I have used selected material in Part 3, including Chapters 13 and 17, in a senior seminar geared
to producing a serious research paper. Along with the basic one-semester course, students who have
been exposed to basic panel data analysis, instrumental variables estimation, and limited dependent
variable models are in a position to read large segments of the applied social sciences literature.
Chapter 17 provides an introduction to the most common limited dependent variable models.
The text is also well suited for an introductory master’s level course, where the emphasis is on
applications rather than on derivations using matrix algebra. Several instructors have used the text to
teach policy analysis at the master’s level. For instructors wanting to present the material in matrix
form, Appendices D and E are self-contained treatments of the matrix algebra and the multiple regression model in matrix form.
At Michigan State, PhD students in many fields that require data analysis—including accounting,
agricultural economics, development economics, economics of education, finance, international economics, labor economics, macroeconomics, political science, and public finance—have found the text
to be a useful bridge between the empirical work that they read and the more theoretical econometrics
they learn at the PhD level.
Suggestions for Designing Your Course Beyond the Basic
I have already commented on the contents of most of the chapters as well as possible outlines for
courses. Here I provide more specific comments about material in chapters that might be covered or
skipped:
Chapter 9 has some interesting examples (such as a wage regression that includes IQ score as
an explanatory variable). The rubric of proxy variables does not have to be formally introduced to
present these kinds of examples, and I typically do so when finishing up cross-sectional analysis. In
Chapter 12, for a one-semester course, I skip the material on serial correlation robust inference for
ordinary least squares as well as dynamic models of heteroskedasticity.
Even in a second course I tend to spend only a little time on Chapter 16, which covers simultaneous equations analysis. I have found that instructors differ widely in their opinions on the importance
of teaching simultaneous equations models to undergraduates. Some think this material is fundamental; others think it is rarely applicable. My own view is that simultaneous equations models are
overused (see Chapter 16 for a discussion). If one reads applications carefully, omitted variables and
measurement error are much more likely to be the reason one adopts instrumental variables estimation,
and this is why I use omitted variables to motivate instrumental variables estimation in Chapter 15.
Still, simultaneous equations models are indispensable for estimating demand and supply functions,
and they apply in some other important cases as well.
Chapter 17 is the only chapter that considers models inherently nonlinear in their parameters,
and this puts an extra burden on the student. The first material one should cover in this chapter is on
probit and logit models for binary response. My presentation of Tobit models and censored regression
still appears to be novel in introductory texts. I explicitly recognize that the Tobit model is applied to
corner solution outcomes on random samples, while censored regression is applied when the data collection process censors the dependent variable at essentially arbitrary thresholds.
Chapter 18 covers some recent important topics from time series econometrics, including testing for unit roots and cointegration. I cover this material only in a second-semester course at either
the undergraduate or master’s level. A fairly detailed introduction to forecasting is also included in
Chapter 18.
58860_fm_hr_i-xxii.indd 15
10/23/18 6:11 PM
xvi
Preface
Chapter 19, which would be added to the syllabus for a course that requires a term paper, is much
more extensive than similar chapters in other texts. It summarizes some of the methods appropriate
for various kinds of problems and data structures, points out potential pitfalls, explains in some detail
how to write a term paper in empirical economics, and includes suggestions for possible projects.
What’s Changed?
I have added new exercises to many chapters, including to the Math Refresher and Advanced
Treatment appendices. Some of the new computer exercises use new data sets, including a data set
on performance of men’s college basketball teams. I have also added more challenging problems that
require derivations.
There are several notable changes to the text. An important organizational change, which should
facilitate a wider variety of teaching tastes, is that the notion of binary, or dummy, explanatory variables is introduced in Chapter 2. There, it is shown that ordinary least squares estimation leads to a
staple in basic statistics: the difference in means between two subgroups in a population. By introducing qualitative factors into regression early on, the instructor is able to use a wider variety of empirical
examples from the very beginning.
The early discussion of binary explanatory variables allows for a formal introduction of potential,
or counterfactual, outcomes, which is indispensable in the modern literature on estimating causal
effects. The counterfactual approach to studying causality appears in previous editions, but Chapters 2,
3, 4, and 7 now explicitly include new sections on the modern approach to causal inference. Because
basic policy analysis involves the binary decision to participate in a program or not, a leading example
of using dummy independent variables in simple and multiple regression is to evaluate policy interventions. At the same time, the new material is incorporated into the text so that instructors not wishing to cover the potential outcomes framework may easily skip the material. Several end-of-chapter
problems concern extensions of the basic potential outcomes framework, which should be valuable
for instructors wishing to cover that material.
Chapter 3 includes a new section on different ways that one can apply multiple regression,
including problems of pure prediction, testing efficient markets, and culminating with a discussion of
estimating treatment or causal effects. I think this section provides a nice way to organize students’
thinking about the scope of multiple regression after they have seen the mechanics of ordinary least
squares (OS) and several examples. As with other new material that touches on causal effects, this
material can be skipped without loss of continuity. A new section in Chapter 7 continues the discussion
of potential outcomes, allowing for nonconstant treatment effects. The material is a nice illustration
of estimating different regression functions for two subgroups from a population. New problems
in this chapter that allow the student more experience in using full regression adjustment to estimate
causal effects.
One notable change to Chapter 9 is a more detailed discussion of using missing data indicators
when data are missing on one or more of the explanatory variables. The assumptions underlying the
method are discussed in more detail than in the previous edition.
Chapter 12 has been reorganized to reflect a more modern treatment of the problem of serial
correlation in the errors of time series regression models. The new structure first covers adjusting the
OLS standard errors to allow general forms of serial correlation. Thus, the chapter outline now parallels that in Chapter 8, with the emphasis in both cases on OLS estimation but making inference robust
to violation of standard assumptions. Correcting for serial correlation using generalized least squares
now comes after OLS and the treatment of testing for serial correlation.
The advanced chapters also include several improvements. Chapter 13 now discusses, at an accessible level, extensions of the standard difference-in-differences setup, allowing for multiple control
58860_fm_hr_i-xxii.indd 16
10/23/18 6:11 PM
Preface
xvii
groups, multiple time periods, and even group-specific trends. In addition, the chapter includes a
more detailed discussion of computing standard errors robust to serial correlation when using firstdifferencing estimation with panel data.
Chapter 14 now provides more detailed discussions of several important issues in estimating
panel data models by fixed effects, random effects, and correlated random effects (CRE). The CRE
approach with missing data is discussed in more detail, as is how one accounts for general functional
forms, such as squares and interactions, which are covered in the cross-sectional setting in Chapter 6.
An expanded section on general policy analysis with panel data should be useful for courses with an
emphasis on program interventions and policy evaluation.
Chapter 16, which still covers simultaneous equations models, now provides an explicit link
between the potential outcomes framework and specification of simultaneous equations models.
Chapter 17 now includes a discussion of using regression adjustment for estimating causal (treatment) effects when the outcome variable has special features, such as when the outcome itself is a
binary variable. Then, as the reader is asked to explore in a new problem, logit and probit models can
be used to obtain more reliable estimates of average treatment effects by estimating separate models
for each treatment group.
Chapter 18 now provides more details about how one can compute a proper standard error for
a forecast (as opposed to a prediction) interval. This should help the advanced reader understand in
more detail the nature of the uncertainty in the forecast.
About MindTap™
MindTap is an outcome-driven application that propels students from memorization to mastery.
It’s the only platform that gives you complete ownership of your course. With it, you can challenge
every student, build their confidence, and empower them to be unstoppable.
Access Everything You Need In One Place. Cut down on prep with preloaded, organized course
materials in MindTap. Teach more efficiently with interactive multimedia, assignments, quizzes and
more. And give your students the power to read, listen and study on their phones, so they can learn on
their terms.
Empower Your Students To Reach Their Potential. Twelve distinct metrics give you actionable
insights into student engagement. Identify topics troubling your entire class and instantly communicate with struggling students. And students can track their scores to stay motivated toward their goals.
Together, you can accelerate progress.
Your Course. Your Content. Only MindTap gives you complete control over your course. You
have the flexibility to reorder textbook chapters, add your own notes and embed a variety of content
including OER. Personalize course content to your students’ needs. They can even read your notes,
add their own and highlight key text to aid their progress.
A Dedicated Team, Whenever You Need Them. MindTap isn’t just a tool; it’s backed by a personalized team eager to support you. Get help setting up your course and tailoring it to your specific
objectives. You’ll be ready to make an impact from day one. And, we’ll be right here to help you and
your students throughout the semester—and beyond.
Design Features
In addition to the didactic material in the chapter, I have included two features to help students
better understand and apply what they are learning. Each chapter contains many numbered examples.
Several of these are case studies drawn from recently published papers. I have used my judgment to
simplify the analysis, hopefully without sacrificing the main point. The “Going Further Questions” in
58860_fm_hr_i-xxii.indd 17
10/23/18 6:11 PM
xviii
Preface
the chapter provide students an opportunity to “go further” in learning the material through analysis
or application. Students will find immediate feedback for these questions in the end of the text.
The end-of-chapter problems and computer exercises are heavily oriented toward empirical work,
rather than complicated derivations. The students are asked to reason carefully based on what they
have learned. The computer exercises often expand on the in-text examples. Several exercises use data
sets from published works or similar data sets that are motivated by published research in economics
and other fields.
A pioneering feature of this introductory econometrics text is the extensive glossary. The short
definitions and descriptions are a helpful refresher for students studying for exams or reading empirical research that uses econometric methods. I have added and updated several entries for the seventh
edition.
Instructional Tools
Cengage offers various supplements for instructors and students who use this book. I would like
to thank the Subject Matter Expert team who worked on these supplements and made teaching and
learning easy.
C. Patrick Scott, Ph.D., Louisiana Tech University (R Videos and Computer exercise reviewer)
Hisham Foad (Aplia Home work reviewer and Glossary)
Kenneth H. Brown, Missouri State University (R Videos creator)
Scott Kostyshak, University of Florida (R Videos reviewer)
Ujwal Kharel (Test Bank and Adaptive Test Prep)
Data Sets—Available in Six Formats
With more than 100 data sets in six different formats, including Stata®, R, EViews®, Minitab®,
Microsoft® Excel, and Text, the instructor has many options for problem sets, examples, and term
projects. Because most of the data sets come from actual research, some are very large. Except for
partial lists of data sets to illustrate the various data structures, the data sets are not reported in the
text. This book is geared to a course where computer work plays an integral role.
Updated Data Sets Handbook
An extensive data description manual is also available online. This manual contains a list of data
sources along with suggestions for ways to use the data sets that are not described in the text. This
unique handbook, created by author Jeffrey M. Wooldridge, lists the source of all data sets for quick
reference and how each might be used. Because the data book contains page numbers, it is easy to
see how the author used the data in the text. Students may want to view the descriptions of each data
set and it can help guide instructors in generating new homework exercises, exam problems, or term
projects. The author also provides suggestions on improving the data sets in this detailed resource that
is available on the book’s companion website at http://login.cengage.com and students can access it
free at www.cengage.com.
58860_fm_hr_i-xxii.indd 18
10/23/18 6:11 PM
Preface
xix
Instructor’s Manual with Solutions
REVISED INSTRUCTOR’S MANUAL WITH SOLUTIONS SAVES TIME IN PREPARATION
AND GRADING. The online Instructor’s Manual with solutions contains answers to all exercises in
this edition. Teaching tips provide suggestions for presenting each chapter’s material. The Instructor’s
Manual also contains sources for each of the data files with suggestions for using the data to develop
problem sets, exams, and term papers. The Instructor’s Manual is password-protected and available
for download on the book’s companion website.
Test Bank
Cengage Testing, powered by Cognero® is a flexible, online system that allows you to import,
edit, and manipulate content from the text’s test bank or elsewhere, including your own favorite test
questions; create multiple test versions in an instant; and deliver tests from your LMS, your classroom, or wherever you want.
PowerPoint Slides
UPDATED POWERPOINT ® SLIDES BRING LECTURES TO LIFE WHILE VISUALLY
CLARIFYING CONCEPTS. Exceptional PowerPoint® presentation slides, created specifically for
this edition, help you create engaging, memorable lectures. The slides are particularly useful for clarifying advanced topics in Part 3. You can modify or customize the slides for your specific course.
PowerPoint® slides are available for convenient download on the instructor-only, password-protected
section of the book’s companion website.
Scientific Word Slides
UPDATED SCIENTIFIC WORD® SLIDES REINFORCE TEXT CONCEPTS AND LECTURE
PRESENTATIONS. Created by the text author, this edition’s Scientific Word ® slides reinforce the
book’s presentation slides while highlighting the benefits of Scientific Word®, the application created by MacKichan software, Inc. for specifically composing mathematical, scientific and technical documents using LaTeX typesetting. These slides are based on the author’s actual lectures and
are available for convenient download on the password-protected section of the book’s companion
website.
Student Supplements
Student Solutions Manual
Now your student’s can maximize their study time and further their course success with this dynamic
online resource. This helpful Solutions Manual includes detailed steps and solutions to odd-numbered
problems as well as computer exercises in the text. This supplement is available as a free resource at
www.cengagebrain.com.
58860_fm_hr_i-xxii.indd 19
10/23/18 6:11 PM
xx
Preface
Acknowledgments
I would like to thank those who reviewed and provided helpful comments for this and previous
­editions of the text:
Erica Johnson, Gonzaga University
Mary Ellen Benedict, Bowling Green
State University
Yanan Di, Wagner College
Chirok Han, Korea University
John Fitzgerald, Bowdoin College
Yan Li, Temple University
Philip N. Jefferson, Swarthmore
College
Melissa Tartari, Yale University
Michael Allgrunn, University of
South Dakota
Yongsheng Wang, Washington and
Jefferson College
Gregory Colman, Pace University
Sheng-Kai Chang, National Taiwan
University
Yoo-Mi Chin, Missouri University of
Science and Technology
Damayanti Ghosh, Binghamton
University
Arsen Melkumian, Western Illinois
University
Susan Averett, Lafayette College
Kevin J. Murphy, Oakland University
Kristine Grimsrud, University of New
Mexico
Will Melick, Kenyon College
Kevin J. Mumford, Purdue
University
Nicolai V. Kuminoff, Arizona State
University
Philip H. Brown, Colby College
Subarna K. Samanta, The College of
New Jersey
Argun Saatcioglu, University of
Kansas
Jing Li, South Dakota State
University
Ken Brown, University of Northern
Iowa
Gary Wagner, University of
Arkansas–Little Rock
Michael R. Jonas, University of San
Francisco
Kelly Cobourn, Boise State University
Melissa Yeoh, Berry College
Timothy Dittmer, Central
Washington University
Nikolaos Papanikolaou, SUNY at
New Paltz
Daniel Fischmar, Westminster
College
Konstantin Golyaev, University of
Minnesota
Subha Mani, Fordham University
Soren Hauge, Ripon College
Kevin Williams, University of
Minnesota
Hailong Qian, Saint Louis
University
Rod Hissong, University of Texas at
Arlington
58860_fm_hr_i-xxii.indd 20
Steven Cuellar, Sonoma State
University
John Maluccio, Middlebury College
James Warner, College of Wooster
Christopher Magee, Bucknell
University
Andrew Ewing, Eckerd College
Debra Israel, Indiana State
University
10/23/18 6:11 PM
xxi
Preface
Jay Goodliffe, Brigham Young
University
Stanley R. Thompson, The Ohio
State University
Michael Robinson, Mount Holyoke
College
Ivan Jeliazkov, University of
California, Irvine
Heather O’Neill, Ursinus College
Leslie Papke, Michigan State
University
Timothy Vogelsang, Michigan State
University
Stephen Woodbury, Michigan State
University
Some of the changes I discussed earlier were driven by comments I received from people on this
list, and I continue to mull over other specific suggestions made by one or more reviewers.
Many students and teaching assistants, too numerous to list, have caught mistakes in earlier
­editions or have suggested rewording some paragraphs. I am grateful to them.
As always, it was a pleasure working with the team at Cengage Learning. Michael Parthenakis,
my longtime Product Manager, has learned very well how to guide me with a firm yet gentle hand.
Anita Verma and Ethan Crist quickly mastered the difficult challenges of being the content and subject matter expert team of a dense, technical textbook. Their careful reading of the manuscript and
fine eye for detail have improved this seventh ­edition considerably.
This book is dedicated to my family: Leslie, Edmund, and R.G.
Jeffrey M. Wooldridge
58860_fm_hr_i-xxii.indd 21
10/23/18 6:11 PM
About the Author
Jeffrey M. Wooldridge is University Distinguished Professor of Economics at Michigan State
University, where he has taught since 1991. From 1986 to 1991, he was an assistant professor of economics at the Massachusetts Institute of Technology. He received his bachelor of arts, with majors in
computer science and economics, from the University of California, Berkeley, in 1982, and received
his doctorate in economics in 1986 from the University of California, San Diego. He has published
more than 60 articles in internationally recognized journals, as well as several book chapters. He
is also the author of Econometric Analysis of Cross Section and Panel Data, second edition. His
awards include an Alfred P. Sloan Research Fellowship, the Plura Scripsit award from Econometric
Theory, the Sir Richard Stone prize from the Journal of Applied Econometrics, and three graduate
teacher-of-the-year awards from MIT. He is a fellow of the Econometric Society and of the Journal
of Econometrics. He is past editor of the Journal of Business and Economic Statistics, and past
econometrics coeditor of Economics Letters. He has served on the editorial boards of Econometric
Theory, the Journal of Economic Literature, the Journal of Econometrics, the Review of Economics
and Statistics, and the Stata Journal. He has also acted as an occasional econometrics consultant for
Arthur Andersen, Charles River Associates, the Washington State Institute for Public Policy, Stratus
Consulting, and Industrial Economics, Incorporated.
xxii
58860_fm_hr_i-xxii.indd 22
10/23/18 6:11 PM
chapter
1
The Nature
of Econometrics
and Economic Data
C
hapter 1 discusses the scope of econometrics and raises general issues that arise in the
application of econometric methods. Section 1-1 provides a brief discussion about the purpose and scope of econometrics and how it fits into economic analysis. Section 1-2 provides
examples of how one can start with an economic theory and build a model that can be estimated
using data. Section 1-3 examines the kinds of data sets that are used in business, economics, and
other social sciences. Section 1-4 provides an intuitive discussion of the difficulties associated with
­inferring ­causality in the social sciences.
1-1 What Is Econometrics?
Imagine that you are hired by your state government to evaluate the effectiveness of a publicly
funded job training program. Suppose this program teaches workers various ways to use computers in
the manufacturing process. The 20-week program offers courses during nonworking hours. Any
hourly manufacturing worker may participate, and enrollment in all or part of the program is voluntary. You are to determine what, if any, effect the training program has on each worker’s subsequent
hourly wage.
Now, suppose you work for an investment bank. You are to study the returns on different investment strategies involving short-term U.S. treasury bills to decide whether they comply with implied
economic theories.
The task of answering such questions may seem daunting at first. At this point, you may only
have a vague idea of the kind of data you would need to collect. By the end of this introductory
econometrics course, you should know how to use econometric methods to formally evaluate a job
training program or to test a simple economic theory.
1
58860_ch01_hr_001-018.indd 1
10/18/18 4:04 PM
2
CHAPTER 1 The Nature of Econometrics and Economic Data
Econometrics is based upon the development of statistical methods for estimating economic
relationships, testing economic theories, and evaluating and implementing government and business
policy. A common application of econometrics is the forecasting of such important macroeconomic
variables as interest rates, inflation rates, and gross domestic product (GDP). Whereas forecasts of
economic indicators are highly visible and often widely published, econometric methods can be used
in economic areas that have nothing to do with macroeconomic forecasting. For example, we will
study the effects of political campaign expenditures on voting outcomes. We will consider the effect
of school spending on student performance in the field of education. In addition, we will learn how to
use econometric methods for forecasting economic time series.
Econometrics has evolved as a separate discipline from mathematical statistics because the former focuses on the problems inherent in collecting and analyzing nonexperimental economic data.
Nonexperimental data are not accumulated through controlled experiments on individuals, firms,
or segments of the economy. (Nonexperimental data are sometimes called observational data, or
retrospective data, to emphasize the fact that the researcher is a passive collector of the data.)
Experimental data are often collected in laboratory environments in the natural sciences, but
they are more difficult to obtain in the social sciences. Although some social experiments can be
devised, it is often impossible, prohibitively expensive, or morally repugnant to conduct the kinds
of controlled experiments that would be needed to address economic issues. We give some specific
examples of the differences between experimental and nonexperimental data in Section 1-4.
Naturally, econometricians have borrowed from mathematical statisticians whenever possible.
The method of multiple regression analysis is the mainstay in both fields, but its focus and interpretation can differ markedly. In addition, economists have devised new techniques to deal with the complexities of economic data and to test the predictions of economic theories.
1-2 Steps in Empirical Economic Analysis
Econometric methods are relevant in virtually every branch of applied economics. They come into
play either when we have an economic theory to test or when we have a relationship in mind that has
some importance for business decisions or policy analysis. An empirical analysis uses data to test a
theory or to estimate a relationship.
How does one go about structuring an empirical economic analysis? It may seem obvious, but
it is worth emphasizing that the first step in any empirical analysis is the careful formulation of the
question of interest. The question might deal with testing a certain aspect of an economic theory, or it
might pertain to testing the effects of a government policy. In principle, econometric methods can be
used to answer a wide range of questions.
In some cases, especially those that involve the testing of economic theories, a formal economic
model is constructed. An economic model consists of mathematical equations that describe various
relationships. Economists are well known for their building of models to describe a vast array of
behaviors. For example, in intermediate microeconomics, individual consumption decisions, subject
to a budget constraint, are described by mathematical models. The basic premise underlying these
models is utility maximization. The assumption that individuals make choices to maximize their
well-being, subject to resource constraints, gives us a very powerful framework for creating tractable
economic models and making clear predictions. In the context of consumption decisions, utility maximization leads to a set of demand equations. In a demand equation, the quantity demanded of each
commodity depends on the price of the goods, the price of substitute and complementary goods, the
consumer’s income, and the individual’s characteristics that affect taste. These equations can form the
basis of an econometric analysis of consumer demand.
Economists have used basic economic tools, such as the utility maximization framework, to
explain behaviors that at first glance may appear to be noneconomic in nature. A classic example is
Becker’s (1968) economic model of criminal behavior.
58860_ch01_hr_001-018.indd 2
10/18/18 4:04 PM
CHAPTER 1 The Nature of Econometrics and Economic Data
Example 1.1
3
Economic Model of Crime
In a seminal article, Nobel Prize winner Gary Becker postulated a utility maximization framework to
describe an individual’s participation in crime. Certain crimes have clear economic rewards, but most
criminal behaviors have costs. The opportunity costs of crime prevent the criminal from participating
in other activities such as legal employment. In addition, there are costs associated with the possibility
of being caught and then, if convicted, the costs associated with incarceration. From Becker’s perspective, the decision to undertake illegal activity is one of resource allocation, with the benefits and
costs of competing activities taken into account.
Under general assumptions, we can derive an equation describing the amount of time spent in
criminal activity as a function of various factors. We might represent such a function as
where
y 5 f 1 x1, x2, x3, x4, x5, x6, x7 2 ,
[1.1]
y 5 hours spent in criminal activities,
x1 5 “wage” for an hour spent in criminal activity,
x2 5 hourly wage in legal employment,
x3 5 income other than from crime or employment,
x4 5 probability of getting caught,
x5 5 probability of being convicted if caught,
x6 5 expected sentence if convicted, and
x7 5 age.
Other factors generally affect a person’s decision to participate in crime, but the list above is representative of what might result from a formal economic analysis. As is common in economic theory,
we have not been specific about the function f(.) in (1.1). This function depends on an underlying utility function, which is rarely known. Nevertheless, we can use economic theory—or introspection—to
predict the effect that each variable would have on criminal activity. This is the basis for an econometric
analysis of individual criminal activity.
Formal economic modeling is sometimes the starting point for empirical analysis, but it is more common to use economic theory less formally, or even to rely entirely on intuition. You may agree that the determinants of criminal behavior appearing in equation (1.1) are reasonable based on common sense; we might
arrive at such an equation directly, without starting from utility maximization. This view has some merit,
although there are cases in which formal derivations provide insights that intuition can overlook.
Next is an example of an equation that we can derive through somewhat informal reasoning.
Example 1.2
Job Training and Worker Productivity
Consider the problem posed at the beginning of Section 1-1. A labor economist would like to examine
the effects of job training on worker productivity. In this case, there is little need for formal economic
theory. Basic economic understanding is sufficient for realizing that factors such as education, experience, and training affect worker productivity. Also, economists are well aware that workers are paid
commensurate with their productivity. This simple reasoning leads to a model such as
where
wage 5 f 1 educ, exper, training 2 ,
[1.2]
wage1 5 hourly wage,
educ1 5 years of formal education,
exper1 5 years of workforce experience, and
training15 weeks spent in job training.
Again, other factors generally affect the wage rate, but equation (1.2) captures the essence of the problem.
58860_ch01_hr_001-018.indd 3
10/18/18 4:04 PM
4
CHAPTER 1 The Nature of Econometrics and Economic Data
After we specify an economic model, we need to turn it into what we call an econometric model.
Because we will deal with econometric models throughout this text, it is important to know how an
econometric model relates to an economic model. Take equation (1.1) as an example. The form of the
function f(.) must be specified before we can undertake an econometric analysis. A second issue concerning (1.1) is how to deal with variables that cannot reasonably be observed. For example, consider
the wage that a person can earn in criminal activity. In principle, such a quantity is well defined, but it
would be difficult if not impossible to observe this wage for a given individual. Even variables such as
the probability of being arrested cannot realistically be obtained for a given individual, but at least we
can observe relevant arrest statistics and derive a variable that approximates the probability of arrest.
Many other factors affect criminal behavior that we cannot even list, let alone observe, but we must
somehow account for them.
The ambiguities inherent in the economic model of crime are resolved by specifying a particular
econometric model:
crime 5 b0 1 b1wage 1 b2othinc 1 b3 freqarr 1 b4 freqconv
1 b5avgsen 1 b6age 1 u,
[1.3]
where
crime 5 some measure of the frequency of criminal activity,
wage
5 the wage that can be earned in legal employment,
othinc 5 the income from other sources (assets, inheritance, and so on),
freqarr 5 the frequency of arrests for prior infractions (to approximate the probability of arrest),
freqconv 5 the frequency of conviction, and
avgsen 5 the average sentence length after conviction.
The choice of these variables is determined by the economic theory as well as data considerations.
The term u contains unobserved factors, such as the wage for criminal activity, moral character, family background, and errors in measuring things like criminal activity and the probability of arrest. We
could add family background variables to the model, such as number of siblings, parents’ education,
and so on, but we can never eliminate u entirely. In fact, dealing with this error term or disturbance
term is perhaps the most important component of any econometric analysis.
The constants b0, b1, . . . , b6 are the parameters of the econometric model, and they describe the
directions and strengths of the relationship between crime and the factors used to determine crime in
the model.
A complete econometric model for Example 1.2 might be
wage 5 b0 1 b1educ 1 b2exper 1 b3training 1 u,
[1.4]
where the term u contains factors such as “innate ability,” quality of education, family background,
and the myriad other factors that can influence a person’s wage. If we are specifically concerned
about the effects of job training, then b3 is the parameter of interest.
For the most part, econometric analysis begins by specifying an econometric model, without consideration of the details of the model’s creation. We generally follow this approach, largely because
careful derivation of something like the economic model of crime is time consuming and can take us
into some specialized and often difficult areas of economic theory. Economic reasoning will play a
role in our examples, and we will merge any underlying economic theory into the econometric model
specification. In the economic model of crime example, we would start with an econometric model
such as (1.3) and use economic reasoning and common sense as guides for choosing the variables.
Although this approach loses some of the richness of economic analysis, it is commonly and effectively applied by careful researchers.
Once an econometric model such as (1.3) or (1.4) has been specified, various hypotheses of
interest can be stated in terms of the unknown parameters. For example, in equation (1.3), we might
hypothesize that wage, the wage that can be earned in legal employment, has no effect on criminal
behavior. In the context of this particular econometric model, the hypothesis is equivalent to b1 5 0.
58860_ch01_hr_001-018.indd 4
10/18/18 4:04 PM
CHAPTER 1 The Nature of Econometrics and Economic Data
5
An empirical analysis, by definition, requires data. After data on the relevant variables have been
collected, econometric methods are used to estimate the parameters in the econometric model and to
formally test hypotheses of interest. In some cases, the econometric model is used to make predictions in either the testing of a theory or the study of a policy’s impact.
Because data collection is so important in empirical work, Section 1-3 will describe the kinds of
data that we are likely to encounter.
1-3 The Structure of Economic Data
Economic data sets come in a variety of types. Whereas some econometric methods can be applied
with little or no modification to many different kinds of data sets, the special features of some data
sets must be accounted for or should be exploited. We next describe the most important data structures
encountered in applied work.
1-3a Cross-Sectional Data
A cross-sectional data set consists of a sample of individuals, households, firms, cities, states, countries,
or a variety of other units, taken at a given point in time. Sometimes, the data on all units do not correspond to precisely the same time period. For example, several families may be surveyed during
different weeks within a year. In a pure cross-sectional analysis, we would ignore any minor timing
differences in collecting the data. If a set of families was surveyed during different weeks of the same
year, we would still view this as a cross-sectional data set.
An important feature of cross-sectional data is that we can often assume that they have been
obtained by random sampling from the underlying population. For example, if we obtain information on wages, education, experience, and other characteristics by randomly drawing 500 people from
the working population, then we have a random sample from the population of all working people.
Random sampling is the sampling scheme covered in introductory statistics courses, and it simplifies
the analysis of cross-sectional data. A review of random sampling is contained in Math Refresher C.
Sometimes, random sampling is not appropriate as an assumption for analyzing cross-sectional
data. For example, suppose we are interested in studying factors that influence the accumulation of
family wealth. We could survey a random sample of families, but some families might refuse to report
their wealth. If, for example, wealthier families are less likely to disclose their wealth, then the resulting sample on wealth is not a random sample from the population of all families. This is an illustration of a sample selection problem, an advanced topic that we will discuss in Chapter 17.
Another violation of random sampling occurs when we sample from units that are large relative to
the population, particularly geographical units. The potential problem in such cases is that the population is not large enough to reasonably assume the observations are independent draws. For example,
if we want to explain new business activity across states as a function of wage rates, energy prices,
corporate and property tax rates, services provided, quality of the workforce, and other state characteristics, it is unlikely that business activities in states near one another are independent. It turns out
that the econometric methods that we discuss do work in such situations, but they sometimes need to
be refined. For the most part, we will ignore the intricacies that arise in analyzing such situations and
treat these problems in a random sampling framework, even when it is not technically correct to do so.
Cross-sectional data are widely used in economics and other social sciences. In economics, the
analysis of cross-sectional data is closely aligned with the applied microeconomics fields, such as
labor economics, state and local public finance, industrial organization, urban economics, demography, and health economics. Data on individuals, households, firms, and cities at a given point in time
are important for testing microeconomic hypotheses and evaluating economic policies.
The cross-sectional data used for econometric analysis can be represented and stored in computers. Table 1.1 contains, in abbreviated form, a cross-sectional data set on 526 working individuals
58860_ch01_hr_001-018.indd 5
10/18/18 4:04 PM
6
CHAPTER 1 The Nature of Econometrics and Economic Data
Table 1.1 A Cross-Sectional Data Set on Wages and Other Individual Characteristics
obsno
wage
educ
1
3.10
2
3.24
3
3.00
4
6.00
5
5.30
.
.
.
exper
female
married
11
2
1
0
12
22
1
1
11
2
0
0
8
44
0
1
12
7
0
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
525
11.56
.
16
5
0
1
526
3.50
14
5
1
0
for the year 1976. (This is a subset of the data in the file WAGE1.) The variables include wage (in
dollars per hour), educ (years of education), exper (years of potential labor force experience), female
(an indicator for gender), and married (marital status). These last two variables are binary (zero-one)
in nature and serve to indicate qualitative features of the individual (the person is female or not; the
person is married or not). We will have much to say about binary variables in Chapter 7 and beyond.
The variable obsno in Table 1.1 is the observation number assigned to each person in the sample.
Unlike the other variables, it is not a characteristic of the individual. All econometrics and statistics
software packages assign an observation number to each data unit. Intuition should tell you that, for
data such as that in Table 1.1, it does not matter which person is labeled as observation 1, which person is called observation 2, and so on. The fact that the ordering of the data does not matter for econometric analysis is a key feature of cross-sectional data sets obtained from random sampling.
Different variables sometimes correspond to different time periods in cross-sectional data sets.
For example, to determine the effects of government policies on long-term economic growth, economists have studied the relationship between growth in real per capita GDP over a certain period (say,
1960 to 1985) and variables determined in part by government policy in 1960 (government consumption as a percentage of GDP and adult secondary education rates). Such a data set might be represented as in Table 1.2, which constitutes part of the data set used in the study of cross-country growth
rates by De Long and Summers (1991).
The variable gpcrgdp represents average growth in real per capita GDP over the period 1960
to 1985. The fact that govcons60 (government consumption as a percentage of GDP) and second60
Table 1.2 A Data Set on Economic Growth Rates and Country Characteristics
58860_ch01_hr_001-018.indd 6
obsno
country
gpcrgdp
govcons60
second60
1
Argentina
2
Austria
0.89
9
32
3.32
16
50
3
4
Belgium
2.56
13
69
Bolivia
1.24
18
12
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
61
Zimbabwe
2.30
17
6
10/18/18 4:04 PM
CHAPTER 1 The Nature of Econometrics and Economic Data
7
(percentage of adult population with a secondary education) correspond to the year 1960, while
gpcrgdp is the average growth over the period from 1960 to 1985, does not lead to any special problems in treating this information as a cross-sectional data set. The observations are listed alphabetically by country, but nothing about this ordering affects any subsequent analysis.
1-3b Time Series Data
A time series data set consists of observations on a variable or several variables over time. Examples
of time series data include stock prices, money supply, consumer price index, GDP, annual homicide
rates, and automobile sales figures. Because past events can influence future events and lags in behavior are prevalent in the social sciences, time is an important dimension in a time series data set. Unlike
the arrangement of cross-sectional data, the chronological ordering of observations in a time series
conveys potentially important information.
A key feature of time series data that makes them more difficult to analyze than cross-sectional
data is that economic observations can rarely, if ever, be assumed to be independent across time. Most
economic and other time series are related, often strongly related, to their recent histories. For example,
knowing something about the GDP from last quarter tells us quite a bit about the likely range of the GDP
during this quarter, because GDP tends to remain fairly stable from one quarter to the next. Although
most econometric procedures can be used with both cross-sectional and time series data, more needs
to be done in specifying econometric models for time series data before standard econometric methods
can be justified. In addition, modifications and embellishments to standard econometric techniques have
been developed to account for and exploit the dependent nature of economic time series and to address
other issues, such as the fact that some economic variables tend to display clear trends over time.
Another feature of time series data that can require special attention is the data frequency
at which the data are collected. In economics, the most common frequencies are daily, weekly,
monthly, quarterly, and annually. Stock prices are recorded at daily intervals (excluding Saturday and
Sunday). The money supply in the U.S. economy is reported weekly. Many macroeconomic series are
tabulated monthly, including inflation and unemployment rates. Other macro series are recorded less
frequently, such as every three months (every quarter). GDP is an important example of a quarterly
series. Other time series, such as infant mortality rates for states in the United States, are available
only on an annual basis.
Many weekly, monthly, and quarterly economic time series display a strong seasonal pattern,
which can be an important factor in a time series analysis. For example, monthly data on housing
starts differ across the months simply due to changing weather conditions. We will learn how to deal
with seasonal time series in Chapter 10.
Table 1.3 contains a time series data set obtained from an article by Castillo-Freeman and
Freeman (1992) on minimum wage effects in Puerto Rico. The earliest year in the data set is the first
Table 1.3 Minimum Wage, Unemployment, and Related Data for Puerto Rico
obsno
58860_ch01_hr_001-018.indd 7
year
avgmin
avgcov
prunemp
prgnp
1
1950
0.20
20.1
15.4
878.7
2
1951
0.21
20.7
16.0
925.0
3
1952
0.23
22.6
14.8
1015.9
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
37
1986
3.35
58.1
18.9
4281.6
.
38
1987
3.35
58.2
16.8
4496.7
10/18/18 4:04 PM
8
CHAPTER 1 The Nature of Econometrics and Economic Data
observation, and the most recent year available is the last observation. When econometric methods are
used to analyze time series data, the data should be stored in chronological order.
The variable avgmin refers to the average minimum wage for the year, avgcov is the average coverage rate (the percentage of workers covered by the minimum wage law), prunemp is the unemployment rate, and prgnp is the gross national product, in millions of 1954 dollars. We will use these data
later in a time series analysis of the effect of the minimum wage on employment.
1-3c Pooled Cross Sections
Some data sets have both cross-sectional and time series features. For example, suppose that two
cross-sectional household surveys are taken in the United States, one in 1985 and one in 1990.
In 1985, a random sample of households is surveyed for variables such as income, savings, family size, and so on. In 1990, a new random sample of households is taken using the same survey
questions. To increase our sample size, we can form a pooled cross section by combining the
two years.
Pooling cross sections from different years is often an effective way of analyzing the effects
of a new government policy. The idea is to collect data from the years before and after a key policy
change. As an example, consider the following data set on housing prices taken in 1993 and 1995,
before and after a reduction in property taxes in 1994. Suppose we have data on 250 houses for 1993
and on 270 houses for 1995. One way to store such a data set is given in Table 1.4.
Observations 1 through 250 correspond to the houses sold in 1993, and observations 251 through
520 correspond to the 270 houses sold in 1995. Although the order in which we store the data turns
out not to be crucial, keeping track of the year for each observation is usually very important. This is
why we enter year as a separate variable.
A pooled cross section is analyzed much like a standard cross section, except that we often need
to account for secular differences in the variables across the time. In fact, in addition to increasing the
sample size, the point of a pooled cross-sectional analysis is often to see how a key relationship has
changed over time.
Table 1.4 Pooled Cross Sections: Two Years of Housing Prices
58860_ch01_hr_001-018.indd 8
obsno
year
hprice
proptax
sqrft
bdrms
bthrms
1
1993
85,500
42
1600
3
2.0
2
1993
67,300
36
1440
3
2.5
3
1993
134,000
38
2000
4
2.5
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
250
1993
243,600
41
2600
4
3.0
251
1995
65,000
16
1250
2
1.0
252
1995
182,400
20
2200
4
2.0
253
1995
97,500
15
1540
3
2.0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
520
1995
16
1100
2
1.5
57,200
10/18/18 4:04 PM
CHAPTER 1 The Nature of Econometrics and Economic Data
9
1-3d Panel or Longitudinal Data
A panel data (or longitudinal data) set consists of a time series for each cross-sectional member
in the data set. As an example, suppose we have wage, education, and employment history for a set
of individuals followed over a 10-year period. Or we might collect information, such as investment
and financial data, about the same set of firms over a five-year time period. Panel data can also be
collected on geographical units. For example, we can collect data for the same set of counties in the
United States on immigration flows, tax rates, wage rates, government expenditures, and so on, for
the years 1980, 1985, and 1990.
The key feature of panel data that distinguishes them from a pooled cross section is that the same
cross-sectional units (individuals, firms, or counties in the preceding examples) are followed over a
given time period. The data in Table 1.4 are not considered a panel data set because the houses sold
are likely to be different in 1993 and 1995; if there are any duplicates, the number is likely to be so
small as to be unimportant. In contrast, Table 1.5 contains a two-year panel data set on crime and
related statistics for 150 cities in the United States.
There are several interesting features in Table 1.5. First, each city has been given a number from
1 through 150. Which city we decide to call city 1, city 2, and so on, is irrelevant. As with a pure cross
section, the ordering in the cross section of a panel data set does not matter. We could use the city
name in place of a number, but it is often useful to have both.
A second point is that the two years of data for city 1 fill the first two rows or observations,
­observations 3 and 4 correspond to city 2, and so on. Because each of the 150 cities has two rows of
data, any econometrics package will view this as 300 observations. This data set can be treated as a
pooled cross section, where the same cities happen to show up in each year. But, as we will see in
Chapters 13 and 14, we can also use the panel structure to analyze questions that cannot be answered
by simply viewing this as a pooled cross section.
In organizing the observations in Table 1.5, we place the two years of data for each city adjacent
to one another, with the first year coming before the second in all cases. For just about every practical purpose, this is the preferred way for ordering panel data sets. Contrast this organization with the
way the pooled cross sections are stored in Table 1.4. In short, the reason for ordering panel data as
in Table 1.5 is that we will need to perform data transformations for each city across the two years.
Because panel data require replication of the same units over time, panel data sets, especially
those on individuals, households, and firms, are more difficult to obtain than pooled cross sections.
Not surprisingly, observing the same units over time leads to several advantages over cross-sectional
data or even pooled cross-sectional data. The benefit that we will focus on in this text is that having
Table 1.5 A Two-Year Panel Data Set on City Crime Statistics
58860_ch01_hr_001-018.indd 9
obsno
city
year
murders
population
unem
police
1
1
1986
5
350,000
8.7
440
2
1
1990
8
359,200
7.2
471
3
2
1986
2
64,300
5.4
75
4
2
1990
1
65,100
5.5
75
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
297
149
1986
10
260,700
9.6
286
298
149
1990
6
245,000
9.8
334
299
150
1986
25
543,000
4.3
520
300
150
1990
32
546,200
5.2
493
10/18/18 4:04 PM
10
CHAPTER 1 The Nature of Econometrics and Economic Data
multiple observations on the same units allows us to control for certain unobserved characteristics
of individuals, firms, and so on. As we will see, the use of more than one observation can facilitate
causal inference in situations where inferring causality would be very difficult if only a single cross
section were available. A second advantage of panel data is that they often allow us to study the
importance of lags in behavior or the result of decision making. This information can be significant
because many economic policies can be expected to have an impact only after some time has passed.
Most books at the undergraduate level do not contain a discussion of econometric methods for
panel data. However, economists now recognize that some questions are difficult, if not impossible,
to answer satisfactorily without panel data. As you will see, we can make considerable progress with
simple panel data analysis, a method that is not much more difficult than dealing with a standard
cross-sectional data set.
1-3e A Comment on Data Structures
Part 1 of this text is concerned with the analysis of cross-sectional data, because this poses the fewest conceptual and technical difficulties. At the same time, it illustrates most of the key themes of econometric
analysis. We will use the methods and insights from cross-sectional analysis in the remainder of the text.
Although the econometric analysis of time series uses many of the same tools as cross-sectional
analysis, it is more complicated because of the trending, highly persistent nature of many economic
time series. Examples that have been traditionally used to illustrate the manner in which econometric
methods can be applied to time series data are now widely believed to be flawed. It makes little sense
to use such examples initially, because this practice will only reinforce poor econometric practice.
Therefore, we will postpone the treatment of time series econometrics until Part 2, when the important issues concerning trends, persistence, dynamics, and seasonality will be introduced.
In Part 3, we will treat pooled cross sections and panel data explicitly. The analysis of independently pooled cross sections and simple panel data analysis are fairly straightforward extensions of
pure cross-sectional analysis. Nevertheless, we will wait until Chapter 13 to deal with these topics.
1-4 Causality, Ceteris Paribus, and Counterfactual Reasoning
In most tests of economic theory, and certainly for evaluating public policy, the economist’s goal is
to infer that one variable (such as education) has a causal effect on another variable (such as worker
productivity). Simply finding an association between two or more variables might be suggestive, but
unless causality can be established, it is rarely compelling.
The notion of ceteris paribus—which means “other (relevant) factors being equal”—plays an
important role in causal analysis. This idea has been implicit in some of our earlier discussion, particularly Examples 1.1 and 1.2, but thus far we have not explicitly mentioned it.
You probably remember from introductory economics that most economic questions are ceteris
paribus by nature. For example, in analyzing consumer demand, we are interested in knowing the
effect of changing the price of a good on its quantity demanded, while holding all other factors—such
as income, prices of other goods, and individual tastes—fixed. If other factors are not held fixed, then
we cannot know the causal effect of a price change on quantity demanded.
Holding other factors fixed is critical for policy analysis as well. In the job training example
(Example 1.2), we might be interested in the effect of another week of job training on wages, with
all other components being equal (in particular, education and experience). If we succeed in holding
all other relevant factors fixed and then find a link between job training and wages, we can conclude
that job training has a causal effect on worker productivity. Although this may seem pretty simple,
even at this early stage it should be clear that, except in very special cases, it will not be possible to
literally hold all else equal. The key question in most empirical studies is: Have enough other factors
been held fixed to make a case for causality? Rarely is an econometric study evaluated without raising
this issue.
58860_ch01_hr_001-018.indd 10
10/18/18 4:04 PM
CHAPTER 1 The Nature of Econometrics and Economic Data
11
In most serious applications, the number of factors that can affect the variable of interest—such
as criminal activity or wages—is immense, and the isolation of any particular variable may seem like
a hopeless effort. However, we will eventually see that, when carefully applied, econometric methods
can simulate a ceteris paribus experiment.
The notion of ceteris paribus also can be described through counterfactual reasoning, which
has become an organizing theme in analyzing various interventions, such as policy changes. The idea
is to imagine an economic unit, such as an individual or a firm, in two or more different states of the
world. For example, consider studying the impact of a job training program on workers’ earnings.
For each worker in the relevant population, we can imagine what his or her subsequent earnings
would be under two states of the world: having participated in the job training program and having
not participated. By considering these counterfactual outcomes (also called potential outcomes),
we easily “hold other factors fixed” because the counterfactual thought experiment applies to each
individual separately. We can then think of causality as meaning that the outcome—in this case, labor
­earnings—in the two states of the world differs for at least some indiviuals. The fact that we will
eventually observe each worker in only one state of the world raises important problems of estimation, but that is a separate issue from the issue of what we mean by causality. We formally introduce
an apparatus for discussing counterfactual outcomes in Chapter 2.
At this point, we cannot yet explain how econometric methods can be used to estimate ceteris
paribus effects, so we will consider some problems that can arise in trying to infer causality in economics. We do not use any equations in this discussion. Instead, in each example, we will discuss
what other factors we would like to hold fixed, and sprinkle in some counterfactual reasoning. For
each example, inferring causality becomes relatively easy if we could conduct an appropriate experiment. Thus, it is useful to describe how such an experiment might be structured, and to observe that,
in most cases, obtaining experimental data is impractical. It is also helpful to think about why the
available data fail to have the important features of an experimental data set.
We rely, for now, on your intuitive understanding of such terms as random, independence, and
correlation, all of which should be familiar from an introductory probability and statistics course.
(These concepts are reviewed in Math Refresher B.) We begin with an example that illustrates some
of these important issues.
Example 1.3
Effects of Fertilizer on Crop Yield
Some early econometric studies [for example, Griliches (1957)] considered the effects of new
fertilizers on crop yields. Suppose the crop under consideration is soybeans. Because fertilizer amount
is only one factor affecting yields—some others include rainfall, quality of land, and presenc­e of
­parasites—this issue must be posed as a ceteris paribus question. One way to determine the causal effect
of fertilizer amount on soybean yield is to conduct an experiment, which might include the ­following
steps. Choose several one-acre plots of land. Apply different amounts of fertilizer to each plot and subsequently measure the yields; this gives us a cross-sectional data set. Then, use statistical methods (to
be introduced in Chapter 2) to measure the association between yields and fertilizer amounts.
As described earlier, this may not seem like a very good experiment because we have said nothing about choosing plots of land that are identical in all respects except for the amount of fertilizer.
In fact, choosing plots of land with this feature is not feasible: some of the factors, such as land
quality, cannot even be fully observed. How do we know the results of this experiment can be used
to measure the ceteris paribus effect of fertilizer? The answer depends on the specifics of how fertilizer amounts are chosen. If the levels of fertilizer are assigned to plots independently of other plot
­features that affect yield—that is, other characteristics of plots are completely ignored when deciding
on ­fertilizer amounts—then we are in business. We will justify this statement in Chapter 2.
The next example is more representative of the difficulties that arise when inferring causality in
applied economics.
58860_ch01_hr_001-018.indd 11
10/18/18 4:04 PM
12
CHAPTER 1 The Nature of Econometrics and Economic Data
Example 1.4
Measuring the Return to Education
Labor economists and policy makers have long been interested in the “return to education.” Somewhat
informally, the question is posed as follows: If a person is chosen from the population and given
another year of education, by how much will his or her wage increase? As with the previous examples, this is a ceteris paribus question, which implies that all other factors are held fixed while another
year of education is given to the person. Notice the element of counterfactual reasoning here: we can
imagine the wage of each individual varying with different levels of education, that is, in different
states of the world. Eventually, we obtain data on each worker in only one state of the world: the
education level they actually wound up with, through perhaps a complicated process of intellectual
ability, motivation for learning, parental input, and societal influences.
We can imagine a social planner designing an experiment to get at this issue, much as the agricultural researcher can design an experiment to estimate fertilizer effects. Assume, for the moment, that
the social planner has the ability to assign any level of education to any person. How would this planner emulate the fertilizer experiment in Example 1.3? The planner would choose a group of people
and randomly assign each person an amount of education; some people are given an eighth-grade
education, some are given a high school education, some are given two years of college, and so on.
Subsequently, the planner measures wages for this group of people (where we assume that each person then works in a job). The people here are like the plots in the fertilizer example, where education
plays the role of fertilizer and wage rate plays the role of soybean yield. As with Example 1.3, if levels
of education are assigned independently of other characteristics that affect productivity (such as experience and innate ability), then an analysis that ignores these other factors will yield useful results.
Again, it will take some effort in Chapter 2 to justify this claim; for now, we state it without support.
Unlike the fertilizer-yield example, the experiment described in Example 1.4 is unfeasible. The
ethical issues, not to mention the economic costs, associated with randomly determining education
levels for a group of individuals are obvious. As a logistical matter, we could not give someone only
an eighth-grade education if he or she already has a college degree.
Even though experimental data cannot be obtained for measuring the return to education, we can
certainly collect nonexperimental data on education levels and wages for a large group by sampling
randomly from the population of working people. Such data are available from a variety of surveys
used in labor economics, but these data sets have a feature that makes it difficult to estimate the
ceteris paribus return to education. People choose their own levels of education; therefore, education
levels are probably not determined independently of all other factors affecting wage. This problem is
a feature shared by most nonexperimental data sets.
One factor that affects wage is experience in the workforce. Because pursuing more education
generally requires postponing entering the workforce, those with more education usually have less
experience. Thus, in a nonexperimental data set on wages and education, education is likely to be negatively associated with a key variable that also affects wage. It is also believed that people with more
innate ability often choose higher levels of education. Because higher ability leads to higher wages,
we again have a correlation between education and a critical factor that affects wage.
The omitted factors of experience and ability in the wage example have analogs in the fertilizer
example. Experience is generally easy to measure and therefore is similar to a variable such as rainfall. Ability, on the other hand, is nebulous and difficult to quantify; it is similar to land quality in the
fertilizer example. As we will see throughout this text, accounting for other observed factors, such as
experience, when estimating the ceteris paribus effect of another variable, such as education, is relatively straightforward. We will also find that accounting for inherently unobservable factors, such as
ability, is much more problematic. It is fair to say that many of the advances in econometric methods
have tried to deal with unobserved factors in econometric models.
One final parallel can be drawn between Examples 1.3 and 1.4. Suppose that in the fertilizer
example, the fertilizer amounts were not entirely determined at random. Instead, the assistant who
58860_ch01_hr_001-018.indd 12
10/18/18 4:04 PM
CHAPTER 1 The Nature of Econometrics and Economic Data
13
chose the fertilizer levels thought it would be better to put more fertilizer on the higher-quality plots
of land. (Agricultural researchers should have a rough idea about which plots of land are of better quality, even though they may not be able to fully quantify the differences.) This situation is
completely analogous to the level of schooling being related to unobserved ability in Example 1.4.
Because better land leads to higher yields, and more fertilizer was used on the better plots, any
observed relationship between yield and fertilizer might be spurious.
Difficulty in inferring causality can also arise when studying data at fairly high levels of aggregation,
as the next example on city crime rates shows.
Example 1.5
The Effect of Law Enforcement on City Crime Levels
The issue of how best to prevent crime has been, and will probably continue to be, with us for some
time. One especially important question in this regard is: Does the presence of more police officers on
the street deter crime?
The ceteris paribus question is easy to state: If a city is randomly chosen and given, say, ten
additional police officers, by how much would its crime rates fall? Closely related to this thought
experiment is explicitly setting up counterfactual outcomes: For a given city, what would its crime
rate be under varying sizes of the police force? Another way to state the question is: If two cities are
the same in all respects, except that city A has ten more police officers than city B, by how much
would the two cities’ crime rates differ?
It would be virtually impossible to find pairs of communities identical in all respects except for
the size of their police force. Fortunately, econometric analysis does not require this. What we do need
to know is whether the data we can collect on community crime levels and the size of the police force
can be viewed as experimental. We can certainly imagine a true experiment involving a large collection of cities where we dictate how many police officers each city will use for the upcoming year.
Although policies can be used to affect the size of police forces, we clearly cannot tell each city
how many police officers it can hire. If, as is likely, a city’s decision on how many police officers to hire
is correlated with other city factors that affect crime, then the data must be viewed as nonexperimental.
In fact, one way to view this problem is to see that a city’s choice of police force size and the amount of
crime are simultaneously determined. We will explicitly address such problems in Chapter 16.
The first three examples we have discussed have dealt with cross-sectional data at various levels
of aggregation (for example, at the individual or city levels). The same hurdles arise when inferring
causality in time series problems.
Example 1.6
The Effect of the Minimum Wage on Unemployment
An important, and perhaps contentious, policy issue concerns the effect of the minimum wage on
unemployment rates for various groups of workers. Although this problem can be studied in a variety
of data settings (cross-sectional, time series, or panel data), time series data are often used to look at
aggregate effects. An example of a time series data set on unemployment rates and minimum wages
was given in Table 1.3.
Standard supply and demand analysis implies that, as the minimum wage is increased above
the market clearing wage, we slide up the demand curve for labor and total employment decreases.
(Labor supply exceeds labor demand.) To quantify this effect, we can study the relationship between
employment and the minimum wage over time. In addition to some special difficulties that can arise
in dealing with time series data, there are possible problems with inferring causality. The minimum
wage in the United States is not determined in a vacuum. Various economic and political forces
impinge on the final minimum wage for any given year. (The minimum wage, once determined, is
usually in place for several years, unless it is indexed for inflation.) Thus, it is probable that the
amount of the minimum wage is related to other factors that have an effect on employment levels.
58860_ch01_hr_001-018.indd 13
10/18/18 4:04 PM
14
CHAPTER 1 The Nature of Econometrics and Economic Data
We can imagine the U.S. government conducting an experiment to determine the employment
effects of the minimum wage (as opposed to worrying about the welfare of low-wage workers). The
minimum wage could be randomly set by the government each year, and then the employment outcomes could be tabulated. The resulting experimental time series data could then be analyzed using
fairly simple econometric methods. But this scenario hardly describes how minimum wages are set.
If we can control enough other factors relating to employment, then we can still hope to estimate
the ceteris paribus effect of the minimum wage on employment. In this sense, the problem is very
similar to the previous cross-sectional examples.
Even when economic theories are not most naturally described in terms of causality, they often
have predictions that can be tested using econometric methods. The following example demonstrates
this approach.
Example 1.7
The Expectations Hypothesis
The expectations hypothesis from financial economics states that, given all information available
to investors at the time of investing, the expected return on any two investments is the same. For
­example, consider two possible investments with a three-month investment horizon, purchased at the
same time: (1) Buy a three-month T-bill with a face value of $10,000, for a price below $10,000; in
three months, you receive $10,000. (2) Buy a six-month T-bill (at a price below $10,000) and, in three
months, sell it as a three-month T-bill. Each investment requires roughly the same amount of initial
capital, but there is an important difference. For the first investment, you know exactly what the return
is at the time of purchase because you know the initial price of the three-month T-bill, along with its
face value. This is not true for the second investment: although you know the price of a six-month
T-bill when you purchase it, you do not know the price you can sell it for in three months. Therefore,
there is uncertainty in this investment for someone who has a three-month investment horizon.
The actual returns on these two investments will usually be different. According to the expectations hypothesis, the expected return from the second investment, given all information at the time of
investment, should equal the return from purchasing a three-month T-bill. This theory turns out to be
fairly easy to test, as we will see in Chapter 11.
Summary
In this introductory chapter, we have discussed the purpose and scope of econometric analysis. Econometrics
is used in all applied economics fields to test economic theories, to inform government and private policy
makers, and to predict economic time series. Sometimes, an econometric model is derived from a formal
economic model, but in other cases, econometric models are based on informal economic reasoning and
intuition. The goals of any econometric analysis are to estimate the parameters in the model and to test
hypotheses about these parameters; the values and signs of the parameters determine the validity of an
economic theory and the effects of certain policies.
Cross-sectional, time series, pooled cross-sectional, and panel data are the most common types of data
structures that are used in applied econometrics. Data sets involving a time dimension, such as time series and
panel data, require special treatment because of the correlation across time of most economic time series. Other
issues, such as trends and seasonality, arise in the analysis of time series data but not cross-sectional data.
In Section 1-4, we discussed the notions of causality, ceteris paribus, and counterfactuals. In most
cases, hypotheses in the social sciences are ceteris paribus in nature: all other relevant factors must be fixed
when studying the relationship between two variables. As we discussed, one way to think of the ceteris
paribus requirement is to undertake a thought experiment where the same economic unit operates in different states of the world, such as different policy regimes. Because of the nonexperimental nature of most
data collected in the social sciences, uncovering causal relationships is very challenging.
58860_ch01_hr_001-018.indd 14
10/18/18 4:04 PM
CHAPTER 1 The Nature of Econometrics and Economic Data
15
Key Terms
Causal Effect
Ceteris Paribus
Counterfactual Outcomes
Counterfactual Reasoning
Cross-Sectional Data Set
Data Frequency
Econometric Model
Economic Model
Empirical Analysis
Experimental Data
Nonexperimental Data
Observational Data
Panel Data
Pooled Cross Section
Random Sampling
Retrospective Data
Time Series Data
Problems
1 Suppose that you are asked to conduct a study to determine whether smaller class sizes lead to
improved student performance of fourth graders.
(i) If you could conduct any experiment you want, what would you do? Be specific.
(ii) More realistically, suppose you can collect observational data on several thousand fourth graders in a given state. You can obtain the size of their fourth-grade class and a standardized test
score taken at the end of fourth grade. Why might you expect a negative correlation between
class size and test score?
(iii) Would a negative correlation necessarily show that smaller class sizes cause better
performance? Explain.
2 A justification for job training programs is that they improve worker productivity. Suppose that you
are asked to evaluate whether more job training makes workers more productive. However, rather than
having data on individual workers, you have access to data on manufacturing firms in Ohio. In particular, for each firm, you have information on hours of job training per worker (training) and number of
nondefective items produced per worker hour (output).
(i) Carefully state the ceteris paribus thought experiment underlying this policy question.
(ii) Does it seem likely that a firm’s decision to train its workers will be independent of worker
characteristics? What are some of those measurable and unmeasurable worker characteristics?
(iii) Name a factor other than worker characteristics that can affect worker productivity.
(iv) If you find a positive correlation between output and training, would you have convincingly
established that job training makes workers more productive? Explain.
3 Suppose at your university you are asked to find the relationship between weekly hours spent studying (study) and weekly hours spent working (work). Does it make sense to characterize the problem as
inferring whether study “causes” work or work “causes” study? Explain.
4 States (and provinces) that have control over taxation sometimes reduce taxes in an attempt to spur
economic growth. Suppose that you are hired by a state to estimate the effect of corporate tax rates on,
say, the growth in per capita gross state product (GSP).
(i) What kind of data would you need to collect to undertake a statistical analysis?
(ii) Is it feasible to do a controlled experiment? What would be required?
(iii) Is a correlation analysis between GSP growth and tax rates likely to be convincing? Explain.
Computer Exercises
C1 Use the data in WAGE1 for this exercise.
(i) Find the average education level in the sample. What are the lowest and highest years of education?
(ii) Find the average hourly wage in the sample. Does it seem high or low?
(iii) The wage data are reported in 1976 dollars. Using the Internet or a printed source, find the
Consumer Price Index (CPI) for the years 1976 and 2013.
58860_ch01_hr_001-018.indd 15
10/18/18 4:04 PM
16
CHAPTER 1 The Nature of Econometrics and Economic Data
(iv) Use the CPI values from part (iii) to find the average hourly wage in 2013 dollars. Now does
the average hourly wage seem reasonable?
(v) How many women are in the sample? How many men?
C2 Use the data in BWGHT to answer this question.
(i) How many women are in the sample, and how many report smoking during pregnancy?
(ii) What is the average number of cigarettes smoked per day? Is the average a good measure of the
“typical” woman in this case? Explain.
(iii) Among women who smoked during pregnancy, what is the average number of cigarettes
smoked per day? How does this compare with your answer from part (ii), and why?
(iv) Find the average of fatheduc in the sample. Why are only 1,192 observations used to compute
this average?
(v) Report the average family income and its standard deviation in dollars.
C3 The data in MEAP01 are for the state of Michigan in the year 2001. Use these data to answer the following questions.
(i) Find the largest and smallest values of math4. Does the range make sense? Explain.
(ii) How many schools have a perfect pass rate on the math test? What percentage is this of the
total sample?
(iii) How many schools have math pass rates of exactly 50%?
(iv) Compare the average pass rates for the math and reading scores. Which test is harder to pass?
(v) Find the correlation between math4 and read4. What do you conclude?
(vi) The variable exppp is expenditure per pupil. Find the average of exppp along with its standard
deviation. Would you say there is wide variation in per pupil spending?
(vii) Suppose School A spends $6,000 per student and School B spends $5,500 per student. By what
percentage does School A’s spending exceed School B’s? Compare this to 100 · [log(6,000) –
log(5,500)], which is the approximation percentage difference based on the difference in the
natural logs. (See Section A.4 in Math Refresher A.)
C4 The data in JTRAIN2 come from a job training experiment conducted for low-income men during
1976–1977; see Lalonde (1986).
(i) Use the indicator variable train to determine the fraction of men receiving job training.
(ii) The variable re78 is earnings from 1978, measured in thousands of 1982 dollars. Find the
averages of re78 for the sample of men receiving job training and the sample not receiving job
training. Is the difference economically large?
(iii) The variable unem78 is an indicator of whether a man is unemployed or not in 1978. What
fraction of the men who received job training are unemployed? What about for men who did
not receive job training? Comment on the difference.
(iv) From parts (ii) and (iii), does it appear that the job training program was effective? What would
make our conclusions more convincing?
C5 The data in FERTIL2 were collected on women living in the Republic of Botswana in 1988. The variable children refers to the number of living children. The variable electric is a binary indicator equal to
one if the woman’s home has electricity, and zero if not.
(i) Find the smallest and largest values of children in the sample. What is the average of children?
(ii) What percentage of women have electricity in the home?
(iii) Compute the average of children for those without electricity and do the same for those with
electricity. Comment on what you find.
(iv) From part (iii), can you infer that having electricity “causes” women to have fewer children?
Explain.
58860_ch01_hr_001-018.indd 16
10/18/18 4:04 PM
CHAPTER 1 The Nature of Econometrics and Economic Data
17
C6 Use the data in COUNTYMURDERS to answer this question. Use only the year 1996. The variable
murders is the number of murders reported in the county. The variable execs is the number of executions that took place of people sentenced to death in the given county. Most states in the United States
have the death penalty, but several do not.
(i) How many counties are there in the data set? Of these, how many have zero murders? What
percentage of counties have zero executions? (Remember, use only the 1996 data.)
(ii) What is the largest number of murders? What is the largest number of executions? Compute the
average number of executions and explain why it is so small.
(iii) Compute the correlation coefficient between murders and execs and describe what you find.
(iv) You should have computed a positive correlation in part (iii). Do you think that more executions
cause more murders to occur? What might explain the positive correlation?
C7 The data set in ALCOHOL contains information on a sample of men in the United States. Two key
variables are self-reported employment status and alcohol abuse (along with many other variables).
The variables employ and abuse are both binary, or indicator, variables: they take on only the values
zero and one.
(i) What percentage of the men in the sample report abusing alcohol? What is the employment rate?
(ii) Consider the group of men who abuse alcohol. What is the employment rate?
(iii) What is the employment rate for the group of men who do not abuse alcohol?
(iv) Discuss the difference in your answers to parts (ii) and (iii). Does this allow you to conclude
that alcohol abuse causes unemployment?
C8 The data in ECONMATH were obtained on students from a large university course in introductory
microeconomics. For this problem, we are interested in two variables: score, which is the final course
score, and econhs, which is a binary variable indicating whether a student took an economics course in
high school.
(i) How many students are in the sample? How many students report taking an economics course
in high school?
(ii) Find the average of score for those students who did take a high school economics class. How
does it compare with the average of score for those who did not?
(iii) Do the findings in part (ii) necessarily tell you anything about the causal effect of taking high
school economics on college course peformance? Explain.
(iv) If you want to obtain a good causal estimate of the effect of taking a high school economics
course using the difference in averages, what experiment would you run?
58860_ch01_hr_001-018.indd 17
10/18/18 4:04 PM
58860_ch01_hr_001-018.indd 18
10/18/18 4:04 PM
Part
1
Regression Analysis with
Cross-Sectional Data
P
art 1 of the text covers regression analysis with cross-sectional data. It builds upon a solid
base of college algebra and basic concepts in probability and statistics. Math Refresher A, B,
and C contain complete reviews of these topics.
Chapter 2 begins with the simple linear regression model, where we explain one variable in
terms of another variable. Although simple regression is not widely used in applied econometrics, it
is used occasionally and serves as a natural starting point because the algebra and interpretations
are relatively straightforward.
Chapters 3 and 4 cover the fundamentals of multiple regression analysis, where we allow more
than one variable to affect the variable we are trying to explain. Multiple regression is still the most
commonly used method in empirical research, and so these chapters deserve careful attention.
Chapter 3 focuses on the algebra of the method of ordinary least squares (OLS), while also establishing conditions under which the OLS estimator is unbiased and best linear unbiased. Chapter 4
covers the important topic of statistical inference.
Chapter 5 discusses the large sample, or asymptotic, properties of the OLS estimators. This
provides justification of the inference procedures in Chapter 4 when the errors in a regression
model are not normally distributed. Chapter 6 covers some additional topics in regression analysis,
including advanced functional form issues, data scaling, prediction, and goodness-of-fit. Chapter 7
explains how qualitative information can be incorporated into multiple regression models.
Chapter 8 illustrates how to test for and correct the problem of heteroskedasticity, or nonconstant variance, in the error terms. We show how the usual OLS statistics can be adjusted, and we
also present an extension of OLS, known as weighted least squares, which explicitly accounts for
different variances in the errors. Chapter 9 delves further into the very important problem of correlation between the error term and one or more of the explanatory variables. We demonstrate how the
availability of a proxy variable can solve the omitted variables problem. In addition, we establish the
bias and inconsistency in the OLS estimators in the presence of certain kinds of measurement errors
in the variables. Various data problems are also discussed, including the problem of outliers.
19
58860_ch02_hr_019-065.indd 19
10/18/18 4:05 PM
chapter
2
The Simple
Regression Model
T
he simple regression model can be used to study the relationship between two variables. For
reasons we will see, the simple regression model has limitations as a general tool for empirical
analysis. Nevertheless, it is sometimes appropriate as an empirical tool. Learning how to interpret the simple regression model is good practice for studying multiple regression, which we will do
in subsequent chapters.
2-1 Definition of the Simple Regression Model
Much of applied econometric analysis begins with the following premise: y and x are two variables,
representing some population, and we are interested in “explaining y in terms of x,” or in “studying
how y varies with changes in x.” We discussed some examples in Chapter 1, including: y is soybean
crop yield and x is amount of fertilizer; y is hourly wage and x is years of education; and y is a community crime rate and x is number of police officers.
In writing down a model that will “explain y in terms of x,” we must confront three issues. First,
because there is never an exact relationship between two variables, how do we allow for other factors
to affect y? Second, what is the functional relationship between y and x? And third, how can we be
sure we are capturing a ceteris paribus relationship between y and x (if that is a desired goal)?
We can resolve these ambiguities by writing down an equation relating y to x. A simple
equation is
y 5 b0 1 b1x 1 u.
[2.1]
Equation (2.1), which is assumed to hold in the population of interest, defines the simple linear
regression model. It is also called the two-variable linear regression model or bivariate linear
20
58860_ch02_hr_019-065.indd 20
10/18/18 4:05 PM
CHAPTER 2 The Simple Regression Model
21
Table 2.1 Terminology for Simple Regression
Y
X
Dependent variable
Independent variable
Explained variable
Explanatory variable
Response variable
Control variable
Predicted variable
Predictor variable
Regressand
Regressor
regression model because it relates the two variables x and y. We now discuss the meaning of each
of the quantities in equation (2.1). [Incidentally, the term “regression” has origins that are not especially important for most modern econometric applications, so we will not explain it here. See Stigler
(1986) for an engaging history of regression analysis.]
When related by equation (2.1), the variables y and x have several different names used interchangeably, as follows: y is called the dependent variable, the explained variable, the response
variable, the predicted variable, or the regressand; x is called the independent variable, the
explanatory variable, the control variable, the predictor variable, or the regressor. (The term
covariate is also used for x.) The terms “dependent variable” and “independent variable” are frequently used in econometrics. But be aware that the label “independent” here does not refer to the
statistical notion of independence between random variables (see Math Refresher B).
The terms “explained” and “explanatory” variables are probably the most descriptive. “Response”
and “control” are used mostly in the experimental sciences, where the variable x is under the experimenter’s control. We will not use the terms “predicted variable” and “predictor,” although you sometimes see these in applications that are purely about prediction and not causality. Our terminology for
simple regression is summarized in Table 2.1.
The variable u, called the error term or disturbance in the relationship, represents factors other
than x that affect y. A simple regression analysis effectively treats all factors affecting y other than x as
being unobserved. You can usefully think of u as standing for “unobserved.”
Equation (2.1) also addresses the issue of the functional relationship between y and x. If the other
factors in u are held fixed, so that the change in u is zero, Du 5 0, then x has a linear effect on y:
Dy 5 b1Dx if Du 5 0.
[2.2]
Thus, the change in y is simply b1 multiplied by the change in x. This means that b1 is the slope
parameter in the relationship between y and x, holding the other factors in u fixed; it is of primary
interest in applied economics. The intercept parameter b0, sometimes called the constant term, also
has its uses, although it is rarely central to an analysis.
Example 2.1
Soybean Yield and Fertilizer
Suppose that soybean yield is determined by the model
yield 5 b0 1 b1 fertilizer 1 u,
[2.3]
so that y 5 yield and x 5 fertilizer. The agricultural researcher is interested in the effect of fertilizer
on yield, holding other factors fixed. This effect is given by b1. The error term u contains factors such
as land quality, rainfall, and so on. The coefficient b1 measures the effect of fertilizer on yield, holding other factors fixed: Dyield 5 b1D fertilizer.
58860_ch02_hr_019-065.indd 21
10/18/18 4:05 PM
22
PART 1 Regression Analysis with Cross-Sectional Data
Example 2.2
A Simple Wage Equation
A model relating a person’s wage to observed education and other unobserved factors is
wage 5 b0 1 b1educ 1 u.
[2.4]
If wage is measured in dollars per hour and educ is years of education, then b1 measures the change
in hourly wage given another year of education, holding all other factors fixed. Some of those factors
include labor force experience, innate ability, tenure with current employer, work ethic, and numerous
other things.
The linearity of equation (2.1) implies that a one-unit change in x has the same effect on y,
regardless of the initial value of x. This is unrealistic for many economic applications. For example, in
the wage-education example, we might want to allow for increasing returns: the next year of education has a larger effect on wages than did the previous year. We will see how to allow for such possibilities in Section 2-4.
The most difficult issue to address is whether model (2.1) really allows us to draw ceteris paribus
conclusions about how x affects y. We just saw in equation (2.2) that b1 does measure the effect of
x on y, holding all other factors (in u) fixed. Is this the end of the causality issue? Unfortunately, no.
How can we hope to learn in general about the ceteris paribus effect of x on y, holding other factors
fixed, when we are ignoring all those other factors?
Section 2-5 will show that we are only able to get reliable estimators of b0 and b1 from a random
sample of data when we make an assumption restricting how the unobservable u is related to the
explanatory variable x. Without such a restriction, we will not be able to estimate the ceteris paribus
effect, b1. Because u and x are random variables, we need a concept grounded in probability.
Before we state the key assumption about how x and u are related, we can always make one
assumption about u. As long as the intercept b0 is included in the equation, nothing is lost by assuming that the average value of u in the population is zero. Mathematically,
E 1 u 2 5 0.
[2.5]
Assumption (2.5) says nothing about the relationship between u and x, but simply makes a statement about the distribution of the unobserved factors in the population. Using the previous examples for illustration, we can see that assumption (2.5) is not very restrictive. In Example 2.1, we
lose nothing by normalizing the unobserved factors affecting soybean yield, such as land quality, to
have an average of zero in the population of all cultivated plots. The same is true of the unobserved
factors in Example 2.2. Without loss of generality, we can assume that things such as average
ability are zero in the population of all working people. If you are not convinced, you should work
through Problem 2 to see that we can always redefine the intercept in equation (2.1) to make equation (2.5) true.
We now turn to the crucial assumption regarding how u and x are related. A natural measure of
the association between two random variables is the correlation coefficient. (See Math Refresher B
for definition and properties.) If u and x are uncorrelated, then, as random variables, they are not
linearly related. Assuming that u and x are uncorrelated goes a long way toward defining the sense in
which u and x should be unrelated in equation (2.1). But it does not go far enough, because correlation measures only linear dependence between u and x. Correlation has a somewhat counterintuitive
feature: it is possible for u to be uncorrelated with x while being correlated with functions of x, such
as x2. (See Section B-4 in Math Refresher B for further discussion.) This possibility is not acceptable
for most regression purposes, as it causes problems for interpreting the model and for deriving statistical properties. A better assumption involves the expected value of u given x.
Because u and x are random variables, we can define the conditional distribution of u given any
value of x. In particular, for any x, we can obtain the expected (or average) value of u for that slice of
58860_ch02_hr_019-065.indd 22
10/18/18 4:05 PM
CHAPTER 2 The Simple Regression Model
23
the population described by the value of x. The crucial assumption is that the average value of u does
not depend on the value of x. We can write this assumption as
E1u0x2 5 E1u2.
[2.6]
E 1 y 0 x 2 5 b0 1 b1x.
[2.8]
Equation (2.6) says that the average value of the unobservables is the same across all slices of the
population determined by the value of x and that the common average is necessarily equal to the
average of u over the entire population. When assumption (2.6) holds, we say that u is mean independent of x. (Of course, mean independence is implied by full independence between u and x, an
assumption often used in basic probability and statistics.) When we combine mean independence
with assumption (2.5), we obtain the zero conditional mean assumption, E 1 u 0 x 2 5 0. It is critical
to remember that equation (2.6) is the assumption with impact; assumption (2.5) essentially defines
the intercept, b0.
Let us see what equation (2.6) entails in the wage example. To simplify the discussion, assume
that u is the same as innate ability. Then equation (2.6) requires that the average level of ability is the
same, regardless of years of education. For example, if E 1 abil 0 8 2 denotes the average ability for the
group of all people with eight years of education, and E 1 abil 0 16 2 denotes the average ability among
people in the population with sixteen years of education, then equation (2.6) implies that these must
be the same. In fact, the average ability level must be the same for all education levels. If, for example, we think that average ability increases with years of education, then equation (2.6) is false. (This
would happen if, on average, people with more ability choose to become more educated.) As we cannot observe innate ability, we have no way of knowG o i n g F u rt h e r 2 . 1
ing whether or not average ability is the same for all
education levels. But this is an issue that we must
Suppose that a score on a final exam, score,
address before relying on simple regression analysis.
depends on classes attended (attend) and
unobserved factors that affect exam perforIn the fertilizer example, if fertilizer amounts are
mance (such as student ability). Then
chosen independently of other features of the plots,
then equation (2.6) will hold: the average land quality
score 5 b0 1 b1attend 1 u. [2.7]
will not depend on the amount of fertilizer. However,
When would you expect this model to satisfy
if more fertilizer is put on the higher-quality plots of
equation (2.6)?
land, then the expected value of u changes with the
level of fertilizer, and equation (2.6) fails.
The zero conditional mean assumption gives b1 another interpretation that is often useful. Taking
the expected value of equation (2.1) conditional on x and using E 1 u 0 x 2 5 0 gives
Equation (2.8) shows that the population regression function (PRF), E 1 y 0 x 2 , is a linear function of
x. The linearity means that a one-unit increase in x changes the expected value of y by the amount b1.
For any given value of x, the distribution of y is centered about E 1 y 0 x 2 , as illustrated in Figure 2.1.
It is important to understand that equation (2.8) tells us how the average value of y changes
with x; it does not say that y equals b0 1 b1x for all units in the population. For example, suppose
that x is the high school grade point average and y is the college GPA, and we happen to know that
E 1 colGPA 0 hsGPA 2 5 1.5 1 0.5 hsGPA. [Of course, in practice, we never know the population
intercept and slope, but it is useful to pretend momentarily that we do to understand the nature of
­equation (2.8).] This GPA equation tells us the average college GPA among all students who have a
given high school GPA. So suppose that hsGPA 5 3.6. Then the average colGPA for all high school
graduates who attend college with hsGPA 5 3.6 is 1.5 1 0.5 1 3.6 2 5 3.3. We are certainly not saying that every student with hsGPA 5 3.6 will have a 3.3 college GPA; this is clearly false. The PRF
gives us a relationship between the average level of y at different levels of x. Some students with
hsGPA 5 3.6 will have a college GPA higher than 3.3, and some will have a lower college GPA.
Whether the actual colGPA is above or below 3.3 depends on the unobserved factors in u, and those
differ among students even within the slice of the population with hsGPA 5 3.6.
58860_ch02_hr_019-065.indd 23
10/18/18 4:05 PM
24
PART 1 Regression Analysis with Cross-Sectional Data
Figure 2.1 E 1 y 0 x 2 as a linear function of x.
y
E(y x) 5
x1
x2
0
1
x
1
x3
Given the zero conditional mean assumption E 1 u 0 x 2 5 0, it is useful to view equation (2.1) as
breaking y into two components. The piece b0 1 b1x, which represents E 1 y 0 x 2 , is called the systematic part of y—that is, the part of y explained by x—and u is called the unsystematic part, or the part
of y not explained by x. In Chapter 3, when we introduce more than one explanatory variable, we will
discuss how to determine how large the systematic part is relative to the unsystematic part.
In the next section, we will use assumptions (2.5) and (2.6) to motivate estimators of b0 and b1
given a random sample of data. The zero conditional mean assumption also plays a crucial role in the
statistical analysis in Section 2-5.
2-2 Deriving the Ordinary Least Squares Estimates
Now that we have discussed the basic ingredients of the simple regression model, we will address the
important issue of how to estimate the parameters b0 and b1 in equation (2.1). To do this, we need a
sample from the population. Let { 1 xi, yi 2 : i 5 1, . . . , n} denote a random sample of size n from the
population. Because these data come from equation (2.1), we can write
yi 5 b0 1 b1xi 1 ui
[2.9]
for each i. Here, ui is the error term for observation i because it contains all factors affecting yi other
than xi.
As an example, xi might be the annual income and yi the annual savings for family i during a particular year. If we have collected data on 15 families, then n 5 15. A scatterplot of such a data set is
given in Figure 2.2, along with the (necessarily fictitious) population regression function.
We must decide how to use these data to obtain estimates of the intercept and slope in the population regression of savings on income.
58860_ch02_hr_019-065.indd 24
10/18/18 4:05 PM
CHAPTER 2 The Simple Regression Model
25
Figure 2.2 Scatterplot of savings and income for 15 families, and the population regression
E 1 savings 0 income 2 5 b0 1 b1 income.
savings
E(savings income) 5
0
0
1
1
income
income
0
There are several ways to motivate the following estimation procedure. We will use equation (2.5) and an important implication of assumption (2.6): in the population, u is uncorrelated with
x. Therefore, we see that u has zero expected value and that the covariance between x and u is zero:
E1u2 5 0
and
[2.10]
Cov 1 x,u 2 5 E 1 xu 2 5 0,
[2.11]
E 1 y 2 b 0 2 b 1x 2 5 0
[2.12]
where the first equality in equation (2.11) follows from (2.10). (See Section B-4 in Math Refresher B
for the definition and properties of covariance.) In terms of the observable variables x and y and the
unknown parameters b0 and b1, equations (2.10) and (2.11) can be written as
and
E 3 x 1 y 2 b0 2 b1x 2 4 5 0,
[2.13]
respectively. Equations (2.12) and (2.13) imply two restrictions on the joint probability distribution
of (x,y) in the population. Because there are two unknown parameters to estimate, we might hope that
equations (2.12) and (2.13) can be used to obtain good estimators of b0 and b1. In fact, they can be.
Given a sample of data, we choose estimates b^ 0 and b^ 1 to solve the sample counterparts of equations
(2.12) and (2.13):
n 21 a 1 yi 2 b^ 0 2 b^ 1xi 2 5 0
n
i51
and
n 21 a xi 1 yi 2 b^ 0 2 b^ 1xi 2 5 0.
n
i51
[2.14]
[2.15]
This is an example of the method of moments approach to estimation. (See Section C-4 in Math Refresher C
for a discussion of different estimation approaches.) These equations can be solved for b^ 0 and b^ 1.
58860_ch02_hr_019-065.indd 25
10/18/18 4:05 PM
26
PART 1 Regression Analysis with Cross-Sectional Data
Using the basic properties of the summation operator from Math Refresher A, equation (2.14)
can be rewritten as
y 5 b^ 0 1 b^ 1x,
[2.16]
where y 5 n21 g i51 yi is the sample average of the yi and likewise for x. This equation allows us
to write b^ 0 in terms of b^ 1, y, and x:
n
b^ 0 5 y 2 b^ 1x.
[2.17]
Therefore, once we have the slope estimate b^ 1, it is straightforward to obtain the intercept estimate b^ 0,
given y and x.
Dropping the n21 in (2.15) (because it does not affect the solution) and plugging (2.17) into
(2.15) yields
^
^
a xi 3 yi 2 1 y 2 b1x 2 2 b1xi 4 5 0,
n
i51
which, upon rearrangement, gives
^
a xi 1 yi 2 y 2 5 b1 a xi 1 xi 2 x 2 .
n
n
i51
i51
From basic properties of the summation operator [see (A-7) and (A-8) in Math Refresher A],
2
a xi 1 xi 2 x 2 5 a 1 xi 2 x 2 and a xi 1 yi 2 y 2 5 a 1 xi 2 x 2 1 yi 2 y 2 .
n
n
n
n
i51
i51
i51
i51
Therefore, provided that
2
a 1 xi 2 x 2 . 0,
n
[2.18]
i51
the estimated slope is
a 1 xi 2 x 2 1 yi 2 y 2
n
b^ 1 5
i51
2
a 1 xi 2 x 2
n
.
[2.19]
i51
Equation (2.19) is simply the sample covariance between xi and yi divided by the sample variance
of xi. Using simple algebra we can also write b^ 1 as
s^ y
b^ 1 5 r^ xy ? a b,
s^ x
where r^ xy is the sample correlation between xi and yi and s^ x, s^ y denote the sample standard deviations. (See Math Refresher C for definitions of correlation and standard deviation. Dividing all sums
by n 2 1 does not affect the formulas.) An immediate implication is that if xi and yi are positively correlated in the sample then b^ 1 . 0; if xi and yi are negatively correlated then b^ 1 , 0.
Not surprisingly, the formula for b^ 1 in terms of the sample correlation and sample standard deviations is the sample analog of the population relationship
b1 5 rxy ? a
sy
sx
b,
where all quantities are defined for the entire population. Recognition that b1 is just a scaled version
of rxy highlights an important limitation of simple regression when we do not have experimental data:
in effect, simple regression is an analysis of correlation between two variables, and so one must be
careful in inferring causality.
Although the method for obtaining (2.17) and (2.19) is motivated by (2.6), the only assumption
needed to compute the estimates for a particular sample is (2.18). This is hardly an assumption at all:
(2.18) is true, provided the xi in the sample are not all equal to the same value. If (2.18) fails, then
58860_ch02_hr_019-065.indd 26
10/18/18 4:05 PM
CHAPTER 2 The Simple Regression Model
27
Figure 2.3 A scatterplot of wage against education when educi 5 12 for all i.
wage
0
12
educ
we have either been unlucky in obtaining our sample from the population or we have not specified an
interesting problem (x does not vary in the population). For example, if y 5 wage and x 5 educ, then
(2.18) fails only if everyone in the sample has the same amount of education (for example, if everyone
is a high school graduate; see Figure 2.3). If just one person has a different amount of education, then
(2.18) holds, and the estimates can be computed.
The estimates given in (2.17) and (2.19) are called the ordinary least squares (OLS) estimates
of b0 and b1. To justify this name, for any b^ 0 and b^ 1 define a fitted value for y when x 5 xi as
y^ i 5 b^ 0 1 b^ 1xi.
[2.20]
This is the value we predict for y when x 5 xi for the given intercept and slope. There is a fitted value
for each observation in the sample. The residual for observation i is the difference between the actual
yi and its fitted value:
u^ i 5 yi 2 y^ i 5 yi 2 b^ 0 1 b^ 1xi.
[2.21]
Again, there are n such residuals. [These are not the same as the errors in (2.9), a point we return to in
Section 2-5.] The fitted values and residuals are indicated in Figure 2.4.
Now, suppose we choose b^ 0 and b^ 1 to make the sum of squared residuals,
2
2
^
^
a u^ i 5 a 1 yi 2 b0 2 b1xi 2 ,
n
n
i51
i51
[2.22]
as small as possible. The appendix to this chapter shows that the conditions necessary for 1 b^ 0,b^ 1 2 to
minimize (2.22) are given exactly by equations (2.14) and (2.15), without n21. Equations (2.14) and
(2.15) are often called the first order conditions for the OLS estimates, a term that comes from optimization using calculus (see Math Refresher A). From our previous calculations, we know that the
solutions to the OLS first order conditions are given by (2.17) and (2.19). The name “ordinary least
squares” comes from the fact that these estimates minimize the sum of squared residuals.
58860_ch02_hr_019-065.indd 27
10/18/18 4:05 PM
28
PART 1 Regression Analysis with Cross-Sectional Data
Figure 2.4 Fitted values and residuals.
y
yi
ûi 5 residual
yˆ 5 ˆ 0 1 ˆ 1x
yˆi 5 fitted value
yˆ 1
y1
x1
xi
x
When we view ordinary least squares as minimizing the sum of squared residuals, it is natural
to ask: why not minimize some other function of the residuals, such as the absolute values of the
residuals? In fact, as we will discuss in the more advanced Section 9-6, minimizing the sum of the
absolute values of the residuals is sometimes very useful. But it does have some drawbacks. First, we
cannot obtain formulas for the resulting estimators; given a data set, the estimates must be obtained
by numerical optimization routines. As a consequence, the statistical theory for estimators that minimize the sum of the absolute residuals is very complicated. Minimizing other functions of the residuals, say, the sum of the residuals each raised to the fourth power, has similar drawbacks. (We would
never choose our estimates to minimize, say, the sum of the residuals themselves, as residuals large
in magnitude but with opposite signs would tend to cancel out.) With OLS, we will be able to derive
unbiasedness, consistency, and other important statistical properties relatively easily. Plus, as the
motivation in equations (2.12) and (2.13) suggests, and as we will see in Section 2-5, OLS is suited
for estimating the parameters appearing in the conditional mean function (2.8).
Once we have determined the OLS intercept and slope estimates, we form the OLS regression line:
y^ 5 b^ 0 1 b^ 1x,
[2.23]
where it is understood that b^ 0 and b^ 1 have been obtained using equations (2.17) and (2.19). The
­notation y^ , read as “y hat,” emphasizes that the predicted values from equation (2.23) are estimates.
The intercept, b^ 0, is the predicted value of y when x 5 0, although in some cases it will not make
sense to set x 5 0. In those situations, b^ 0 is not, in itself, very interesting. When using (2.23) to compute predicted values of y for various values of x, we must account for the intercept in the calculations. Equation (2.23) is also called the sample regression function (SRF) because it is the estimated
version of the population regression function E 1 y 0 x 2 5 b0 1 b1x. It is important to remember that
the PRF is something fixed, but unknown, in the population. Because the SRF is obtained for a given
sample of data, a new sample will generate a different slope and intercept in equation (2.23).
In most cases, the slope estimate, which we can write as
b^ 1 5 Dy^ /Dx,
58860_ch02_hr_019-065.indd 28
[2.24]
10/18/18 4:05 PM
CHAPTER 2 The Simple Regression Model
29
is of primary interest. It tells us the amount by which y^ changes when x increases by one unit.
Equivalently,
Dy^ 5 b^ 1Dx,
[2.25]
so that given any change in x (whether positive or negative), we can compute the predicted change in y.
We now present several examples of simple regression obtained by using real data. In other
words, we find the intercept and slope estimates with equations (2.17) and (2.19). Because these
examples involve many observations, the calculations were done using an econometrics software
package. At this point, you should be careful not to read too much into these regressions; they are not
necessarily uncovering a causal relationship. We have said nothing so far about the statistical properties of OLS. In Section 2-5, we consider statistical properties after we explicitly impose assumptions
on the population model equation (2.1).
Example 2.3
CEO Salary and Return on Equity
For the population of chief executive officers, let y be annual salary (salary) in thousands of dollars. Thus, y 5 856.3 indicates an annual salary of $856,300, and y 5 1,452.6 indicates a salary of
$1,452,600. Let x be the average return on equity (roe) for the CEO’s firm for the previous three
years. (Return on equity is defined in terms of net income as a percentage of common equity.) For
example, if roe 5 10, then average return on equity is 10%.
To study the relationship between this measure of firm performance and CEO compensation, we
postulate the simple model
salary 5 b0 1 b1roe 1 u.
The slope parameter b1 measures the change in annual salary, in thousands of dollars, when return
on equity increases by one percentage point. Because a higher roe is good for the company, we think
b1 . 0.
The data set CEOSAL1 contains information on 209 CEOs for the year 1990; these data were obtained
from Business Week (5/6/91). In this sample, the average annual salary is $1,281,120, with the smallest and
largest being $223,000 and $14,822,000, respectively. The average return on equity for the years 1988,
1989, and 1990 is 17.18%, with the smallest and largest values being 0.5% and 56.3%, respectively.
Using the data in CEOSAL1, the OLS regression line relating salary to roe is
salary 5 963.191 1 18.501 roe
[2.26]
n 5 209,
where the intercept and slope estimates have been rounded to three decimal places; we use “salary
hat” to indicate that this is an estimated equation. How do we interpret the equation? First, if the return
on equity is zero, roe 5 0, then the predicted salary is the intercept, 963.191, which equals $963,191
because salary is measured in thousands. Next, we can write the predicted change in salary as a function of the change in roe: Dsalary 5 18.501 1 Droe 2 . This means that if the return on equity increases
by one percentage point, Droe 5 1, then salary is predicted to change by about 18.5, or $18,500.
Because (2.26) is a linear equation, this is the estimated change regardless of the initial salary.
We can easily use (2.26) to compare predicted salaries at different values of roe. Suppose
roe 5 30. Then salary 5 963.191 1 18.501 1 30 2 5 1,518,221, which is just over $1.5 million.
However, this does not mean that a particular CEO whose firm had a roe 5 30 earns $1,518,221.
Many other factors affect salary. This is just our prediction from the OLS regression line (2.26). The
estimated line is graphed in Figure 2.5, along with the population regression function E 1 salary 0 roe 2 .
We will never know the PRF, so we cannot tell how close the SRF is to the PRF. Another sample of data
will give a different regression line, which may or may not be closer to the population regression line.
58860_ch02_hr_019-065.indd 29
10/18/18 4:05 PM
30
PART 1 Regression Analysis with Cross-Sectional Data
Figure 2.5 The OLS regression line salary 5 963.191 1 18.501 roe and the (unknown)
population regression function.
salary
salary 5 963.191 1 18.501 roe
E(salary roe) 5
0
1
roe
1
963.191
roe
Example 2.4
Wage and Education
For the population of people in the workforce in 1976, let y 5 wage, where wage is measured in dollars per hour. Thus, for a particular person, if wage 5 6.75, the hourly wage is $6.75. Let x 5 educ
denote years of schooling; for example, educ 5 12 corresponds to a complete high school education.
Because the average wage in the sample is $5.90, the Consumer Price Index indicates that this amount
is equivalent to $24.90 in 2016 dollars.
Using the data in WAGE1 where n 5 526 individuals, we obtain the following OLS regression
line (or sample regression function):
wage 5 20.90 1 0.54 educ
[2.27]
n 5 526.
We must interpret this equation with caution. The intercept of −0.90 literally means that a person
with no education has a predicted hourly wage of −90¢ an hour. This, of course, is silly. It turns out
that only 18 people in the sample of 526 have less than eight years of education. Consequently, it
is not surprising that the regression line does poorly at very low levels of education. For a person
with eight years of education, the predicted wage
is wage 5 20.90 1 0.54 1 8 2 5 3.42, or $3.42 per
G o i n g F u rt h e r 2 . 2
hour (in 1976 dollars).
The estimated wage from (2.27), when
The slope estimate in (2.27) implies that one
educ 5 8, is $3.42 in 1976 dollars. What
more year of education increases hourly wage by
is this value in 2016 dollars? (Hint: You
54 ¢ an hour. Therefore, four more years of educahave enough information in Example 2.4 to
tion increase the predicted wage by 4 1 0.54 2 5 2.16,
answer this question.)
or $2.16 per hour. These are fairly large effects.
58860_ch02_hr_019-065.indd 30
10/18/18 4:05 PM
CHAPTER 2 The Simple Regression Model
31
Because of the linear nature of (2.27), another year of education increases the wage by the same
amount, regardless of the initial level of education. In Section 2-4, we discuss some methods that
­allow for nonconstant marginal effects of our explanatory variables.
Example 2.5
Voting Outcomes and Campaign Expenditures
The file VOTE1 contains data on election outcomes and campaign expenditures for 173 two-party
races for the U.S. House of Representatives in 1988. There are two candidates in each race, A and
B. Let voteA be the percentage of the vote received by Candidate A and shareA be the percentage of
total campaign expenditures accounted for by Candidate A. Many factors other than shareA affect the
election outcome (including the quality of the candidates and possibly the dollar amounts spent by A
and B). Nevertheless, we can estimate a simple regression model to find out whether spending more
relative to one’s challenger implies a higher percentage of the vote.
The estimated equation using the 173 observations is
voteA 5 26.81 1 0.464 shareA
[2.28]
n 5 173.
This means that if Candidate A’s share of spending increases by one percentage point, Candidate A
receives almost one-half a percentage point (0.464) more of the total vote. Whether or not this is a
causal effect is unclear, but it is not unbelievable. If shareA 5 50, voteA is predicted to be about 50,
or half the vote.
G o i n g F u rt h e r 2 . 3
In Example 2.5, what is the predicted vote
for Candidate A if shareA 5 60 (which
means 60%)? Does this answer seem
reasonable?
In some cases, regression analysis is not used to
determine causality but to simply look at whether two
variables are positively or negatively related, much
like a standard correlation analysis. An example of
this occurs in Computer Exercise C3, where you
are asked to use data from Biddle and Hamermesh
(1990) on time spent sleeping and working to investigate the tradeoff between these two factors.
2-2a A Note on Terminology
In most cases, we will indicate the estimation of a relationship through OLS by writing an equation
such as (2.26), (2.27), or (2.28). Sometimes, for the sake of brevity, it is useful to indicate that an OLS
regression has been run without actually writing out the equation. We will often indicate that equation
(2.23) has been obtained by OLS in saying that we run the regression of
y on x,
[2.29]
or simply that we regress y on x. The positions of y and x in (2.29) indicate which is the dependent
variable and which is the independent variable: We always regress the dependent variable on the
independent variable. For specific applications, we replace y and x with their names. Thus, to obtain
(2.26), we regress salary on roe, or to obtain (2.28), we regress voteA on shareA.
When we use such terminology in (2.29), we will always mean that we plan to estimate the
intercept, b^ 0, along with the slope, b^ 1. This case is appropriate for the vast majority of applications.
58860_ch02_hr_019-065.indd 31
10/18/18 4:05 PM
32
PART 1 Regression Analysis with Cross-Sectional Data
Occasionally, we may want to estimate the relationship between y and x assuming that the intercept
is zero (so that x 5 0 implies that y^ 5 0); we cover this case briefly in Section 2-6. Unless explicitly
stated otherwise, we always estimate an intercept along with a slope.
2-3 Properties of OLS on Any Sample of Data
In the previous section, we went through the algebra of deriving the formulas for the OLS intercept
and slope estimates. In this section, we cover some further algebraic properties of the fitted OLS
regression line. The best way to think about these properties is to remember that they hold, by construction, for any sample of data. The harder task—considering the properties of OLS across all possible random samples of data—is postponed until Section 2-5.
Several of the algebraic properties we are going to derive will appear mundane. Nevertheless,
having a grasp of these properties helps us to figure out what happens to the OLS estimates and
related statistics when the data are manipulated in certain ways, such as when the measurement units
of the dependent and independent variables change.
2-3a Fitted Values and Residuals
We assume that the intercept and slope estimates, b^ 0 and b^ 1, have been obtained for the given sample
of data. Given b^ 0 and b^ 1, we can obtain the fitted value y^ i for each observation. [This is given by equation (2.20).] By definition, each fitted value of y^ i is on the OLS regression line. The OLS residual
associated with observation i, u^ i, is the difference between yi and its fitted value, as given in equation
(2.21). If u^ i is positive, the line underpredicts yi; if u^ i is negative, the line overpredicts yi. The ideal
case for observation i is when u^ i 5 0, but in most cases, every residual is not equal to zero. In other
words, none of the data points must actually lie on the OLS line.
Example 2.6
CEO Salary and Return on Equity
Table 2.2 contains a listing of the first 15 observations in the CEO data set, along with the fitted
values, called salaryhat, and the residuals, called uhat.
The first four CEOs have lower salaries than what we predicted from the OLS regression
line (2.26); in other words, given only the firm’s roe, these CEOs make less than what we
predicted. As can be seen from the positive uhat, the fifth CEO makes more than predicted from
the OLS regression line.
2-3b Algebraic Properties of OLS Statistics
There are several useful algebraic properties of OLS estimates and their associated statistics. We now
cover the three most important of these.
(1) The sum, and therefore the sample average of the OLS residuals, is zero. Mathematically,
a u^ i 5 0.
n
i51
[2.30]
This property needs no proof; it follows immediately from the OLS first order condition (2.14), when
we remember that the residuals are defined by u^ i 5 yi 2 b^ 0 2 b^ 1xi. In other words, the OLS estimates
58860_ch02_hr_019-065.indd 32
10/18/18 4:05 PM
CHAPTER 2 The Simple Regression Model
33
Table 2.2 Fitted Values and Residuals for the First 15 CEOs
obsno
roe
salary
salary
u^
1
14.1
1095
1224.058
−129.0581
2
10.9
1001
1164.854
−163.8542
3
23.5
1122
1397.969
−275.9692
4
5.9
578
1072.348
−494.3484
5
13.8
1368
1218.508
149.4923
6
20.0
1145
1333.215
−188.2151
7
16.4
1078
1266.611
−188.6108
8
16.3
1094
1264.761
−170.7606
9
10.5
1237
1157.454
10
26.3
833
1449.773
−616.7726
11
25.9
567
1442.372
−875.3721
12
26.8
933
1459.023
−526.0231
13
14.8
1339
1237.009
101.9911
14
22.3
937
1375.768
−438.7678
15
56.3
2011
2004.808
79.54626
6.191895
b^ 0 and b^ 1 are chosen to make the residuals add up to zero (for any data set). This says nothing about
the residual for any particular observation i.
(2) The sample covariance between the regressors and the OLS residuals is zero. This follows
from the first order condition (2.15), which can be written in terms of the residuals as
a xiu^ i 5 0.
n
i51
[2.31]
The sample average of the OLS residuals is zero, so the left-hand side of (2.31) is proportional to the
sample covariance between xi and u^ i.
(3) The point 1 x,y 2 is always on the OLS regression line. In other words, if we take equation
(2.23) and plug in x for x, then the predicted value is y. This is exactly what equation (2.16) showed us.
Example 2.7
Wage and Education
For the data in WAGE1, the average hourly wage in the sample is 5.90, rounded to two decimal
places, and the average education is 12.56. If we plug educ 5 12.56 into the OLS regression line
(2.27), we get wage 5 20.90 1 0.54 1 12.56 2 5 5.8824, which equals 5.9 when rounded to the first
decimal place. These figures do not exactly agree because we have rounded the average wage and
education, as well as the intercept and slope estimates. If we did not initially round any of the values,
we would get the answers to agree more closely, but to little useful effect.
Writing each yi as its fitted value, plus its residual, provides another way to interpret an OLS
regression. For each i, write
yi 5 y^ i 1 u^ i.
58860_ch02_hr_019-065.indd 33
[2.32]
10/18/18 4:05 PM
34
PART 1 Regression Analysis with Cross-Sectional Data
From property (1), the average of the residuals is zero; equivalently, the sample average of the fitted
values, y^ i, is the same as the sample average of the yi, or y^ 5 y. Further, properties (1) and (2) can be
used to show that the sample covariance between y^ i and u^ i is zero. Thus, we can view OLS as decomposing each yi into two parts, a fitted value and a residual. The fitted values and residuals are uncorrelated in the sample.
Define the total sum of squares (SST), the explained sum of squares (SSE), and the residual
sum of squares (SSR) (also known as the sum of squared residuals), as follows:
SST ; a 1 yi 2 y 2 2.
n
[2.33]
i51
SSE ; a 1 y^ i 2 y 2 2.
n
[2.34]
i51
SSR ; a u^ 2i .
n
[2.35]
i51
SST is a measure of the total sample variation in the yi; that is, it measures how spread out the yi are in
the sample. If we divide SST by n 2 1, we obtain the sample variance of y, as discussed in Math Refresher C.
Similarly, SSE measures the sample variation in the y^ i (where we use the fact that y^ 5 y), and SSR
measures the sample variation in the u^ i. The total variation in y can always be expressed as the sum of
the explained variation and the unexplained variation SSR. Thus,
[2.36]
SST 5 SSE 1 SSR.
Proving (2.36) is not difficult, but it requires us to use all of the properties of the summation operator
covered in Math Refresher A. Write
2
2
a 1 yi 2 y 2 5 a 3 1 yi 2 y^ i 2 1 1 y^ i 2 y 2 4
n
n
i51
i51
5 a 3 u^ i 1 1 y^ i 2 y 2 4 2
n
i51
5 a u^ 2i 1 2 a u^ i 1 y^ i 2 y 2 1 a 1 y^ i 2 y 2 2
n
n
n
i51
i51
i51
5 SSR 1 2 a u^ i 1 y^ i 2 y 2 1 SSE.
n
i51
Now, (2.36) holds if we show that
a u^ i 1 y^ i 2 y 2 5 0.
n
i51
[2.37]
But we have already claimed that the sample covariance between the residuals and the fitted values is
zero, and this covariance is just (2.37) divided by n 2 1. Thus, we have established (2.36).
Some words of caution about SST, SSE, and SSR are in order. There is no uniform agreement on the names or abbreviations for the three quantities defined in equations (2.33), (2.34),
and (2.35). The total sum of squares is called either SST or TSS, so there is little confusion here.
Unfortunately, the explained sum of squares is sometimes called the “regression sum of squares.”
If this term is given its natural abbreviation, it can easily be confused with the term “residual sum
of squares.” Some regression packages refer to the explained sum of squares as the “model sum of
squares.”
To make matters even worse, the residual sum of squares is often called the “error sum of
squares.” This is especially unfortunate because, as we will see in Section 2-5, the errors and the
residuals are different quantities. Thus, we will always call (2.35) the residual sum of squares or the
sum of squared residuals. We prefer to use the abbreviation SSR to denote the sum of squared residuals, because it is more common in econometric packages.
58860_ch02_hr_019-065.indd 34
10/18/18 4:05 PM
CHAPTER 2 The Simple Regression Model
35
2-3c Goodness-of-Fit
So far, we have no way of measuring how well the explanatory or independent variable, x, explains
the dependent variable, y. It is often useful to compute a number that summarizes how well the OLS
regression line fits the data. In the following discussion, be sure to remember that we assume that an
intercept is estimated along with the slope.
Assuming that the total sum of squares, SST, is not equal to zero—which is true except in
the very unlikely event that all the yi equal the same value—we can divide (2.36) by SST to get
1 5 SSE/SST 1 SSR/SST. The R-squared of the regression, sometimes called the coefficient of
determination, is defined as
R2 ; SSE/SST 5 1 2 SSR/SST.
[2.38]
2
R is the ratio of the explained variation compared to the total variation; thus, it is interpreted as the
fraction of the sample variation in y that is explained by x. The second equality in (2.38) provides
another way for computing R2.
From (2.36), the value of R2 is always between zero and one, because SSE can be no greater than
SST. When interpreting R2, we usually multiply it by 100 to change it into a percent: 100 # R2 is the
percentage of the sample variation in y that is explained by x.
If the data points all lie on the same line, OLS provides a perfect fit to the data. In this case,
R2 5 1. A value of R2 that is nearly equal to zero indicates a poor fit of the OLS line: very little of
the variation in the yi is captured by the variation in the y^ i (which all lie on the OLS regression line).
In fact, it can be shown that R2 is equal to the square of the sample correlation coefficient between yi
and y^ i. This is where the term “R-squared” came from. (The letter R was traditionally used to denote
an estimate of a population correlation coefficient, and its usage has survived in regression analysis.)
Example 2.8
CEO Salary and Return on Equity
In the CEO salary regression, we obtain the following:
salary 5 963.191 1 18.501 roe
[2.39]
n 5 209, R2 5 0.0132.
We have reproduced the OLS regression line and the number of observations for clarity. Using the
R-squared (rounded to four decimal places) reported for this equation, we can see how much of
the variation in salary is actually explained by the return on equity. The answer is: not much. The
firm’s return on equity explains only about 1.3% of the variation in salaries for this sample of 209
CEOs. That means that 98.7% of the salary variations for these CEOs is left unexplained! This lack
of e­ xplanatory power may not be too surprising because many other characteristics of both the firm
and the individual CEO should influence salary; these factors are necessarily included in the errors in
a simple regression analysis.
In the social sciences, low R-squareds in regression equations are not uncommon, especially
for cross-sectional analysis. We will discuss this issue more generally under multiple regression analysis, but it is worth emphasizing now that a seemingly low R-squared does not necessarily mean that an OLS regression equation is useless. It is still possible that (2.39) is a good
estimate of the ceteris paribus relationship between salary and roe; whether or not this is true
does not depend directly on the size of R-squared. Students who are first learning econometrics
tend to put too much weight on the size of the R-squared in evaluating regression equations. For
now, be aware that using R-squared as the main gauge of success for an econometric analysis
can lead to trouble.
58860_ch02_hr_019-065.indd 35
10/18/18 4:05 PM
36
PART 1 Regression Analysis with Cross-Sectional Data
Sometimes, the explanatory variable explains a substantial part of the sample variation in the
dependent variable.
Example 2.9
Voting Outcomes and Campaign Expenditures
In the voting outcome equation in (2.28), R2 5 0.856. Thus, the share of campaign expenditures
explains over 85% of the variation in the election outcomes for this sample. This is a sizable portion.
2-4 Units of Measurement and Functional Form
Two important issues in applied economics are (1) understanding how changing the units of measurement of the dependent and/or independent variables affects OLS estimates and (2) knowing how to
incorporate popular functional forms used in economics into regression analysis. The mathematics
needed for a full understanding of functional form issues is reviewed in Math Refresher A.
2-4a The Effects of Changing Units of Measurement
on OLS Statistics
In Example 2.3, we chose to measure annual salary in thousands of dollars, and the return on equity
was measured as a percentage (rather than as a decimal). It is crucial to know how salary and roe are
measured in this example in order to make sense of the estimates in equation (2.39).
We must also know that OLS estimates change in entirely expected ways when the units
of measurement of the dependent and independent variables change. In Example 2.3, suppose
that, rather than measuring salary in thousands of dollars, we measure it in dollars. Let salardol
be ­salary in dollars (salardol 5 845,761 would be interpreted as $845,761). Of course, salardol
has a simple relationship to the salary measured in thousands of dollars: salardol 5 1,000 # salary.
We do not need to actually run the regression of salardol on roe to know that the estimated
­equation is:
salardol 5 963,191 1 18,501 roe.
[2.40]
We obtain the intercept and slope in (2.40) simply by multiplying the intercept and the slope in (2.39)
by 1,000. This gives equations (2.39) and (2.40) the same interpretation. Looking at (2.40), if roe 5 0,
then salardol 5 963,191, so the predicted salary is $963,191 [the same value we obtained from equation (2.39)]. Furthermore, if roe increases by one, then the predicted salary increases by $18,501;
again, this is what we concluded from our earlier analysis of equation (2.39).
Generally, it is easy to figure out what happens to the intercept and slope estimates when
the dependent variable changes units of measurement. If the dependent variable is multiplied
by the constant c—which means each value in the sample is multiplied by c—then the OLS
intercept and slope estimates are also multiplied by c. (This assumes nothing has changed
about the independent variable.) In the CEO salary example, c 5 1,000 in moving from salary to
salardol.
G o i n g F u rt h e r 2 . 4
We can also use the CEO salary example to
Suppose that salary is measured in hunsee what happens when we change the units of
dreds of dollars, rather than in thousands of
measurement of the independent variable. Define
dollars, say, salarhun. What will be the OLS
roedec 5 roe/100 to be the decimal equivalent of
intercept and slope estimates in the regresroe; thus, roedec 5 0.23 means a return on equity of
sion of salarhun on roe?
23%. To focus on changing the units of measurement
58860_ch02_hr_019-065.indd 36
10/18/18 4:05 PM
CHAPTER 2 The Simple Regression Model
37
of the independent variable, we return to our original dependent variable, salary, which is measured
in thousands of dollars. When we regress salary on roedec, we obtain
salary 5 963.191 1 1,850.1 roedec.
[2.41]
The coefficient on roedec is 100 times the coefficient on roe in (2.39). This is as it should
be. Changing roe by one percentage point is equivalent to Droedec 5 0.01. From (2.41), if
Droedec 5 0.01, then Dsalary 5 1,850.1 1 0.01 2 5 18.501, which is what is obtained by using (2.39).
Note that, in moving from (2.39) to (2.41), the independent variable was divided by 100, and so the
OLS slope estimate was multiplied by 100, preserving the interpretation of the equation. Generally,
if the independent variable is divided or multiplied by some nonzero constant, c, then the OLS slope
coefficient is multiplied or divided by c, respectively.
The intercept has not changed in (2.41) because roedec 5 0 still corresponds to a zero return on
equity. In general, changing the units of measurement of only the independent variable does not affect
the intercept.
In the previous section, we defined R-squared as a goodness-of-fit measure for OLS r­ egression.
We can also ask what happens to R2 when the unit of measurement of either the independent or
the dependent variable changes. Without doing any algebra, we should know the result: the
goodness-of-fit of the model should not depend on the units of measurement of our variables. For
example, the amount of variation in salary explained by the return on equity should not depend on
whether salary is measured in dollars or in thousands of dollars or on whether return on equity is a
percentage or a decimal. This intuition can be verified mathematically: using the definition of R2, it
can be shown that R2 is, in fact, invariant to changes in the units of y or x.
2-4b Incorporating Nonlinearities in Simple Regression
So far, we have focused on linear relationships between the dependent and independent variables.
As we mentioned in Chapter 1, linear relationships are not nearly general enough for all economic
applications. Fortunately, it is rather easy to incorporate many nonlinearities into simple regression
analysis by appropriately defining the dependent and independent variables. Here, we will cover two
possibilities that often appear in applied work.
In reading applied work in the social sciences, you will often encounter regression equations
where the dependent variable appears in logarithmic form. Why is this done? Recall the wage-­
education example, where we regressed hourly wage on years of education. We obtained a slope estimate of 0.54 [see equation (2.27)], which means that each additional year of education is predicted to
increase hourly wage by 54 cents. Because of the linear nature of (2.27), 54 cents is the increase for
either the first year of education or the twentieth year; this may not be reasonable.
Probably a better characterization of how wage changes with education is that each year of education increases wage by a constant percentage. For example, an increase in education from 5 years to
6 years increases wage by, say, 8% (ceteris paribus), and an increase in education from 11 to 12 years
also increases wage by 8%. A model that gives (approximately) a constant percentage effect is
log 1 wage 2 5 b0 1 b1educ 1 u,
[2.42]
%Dwage < 1 100 ? b1 2 Deduc.
[2.43]
where log 1 # 2 denotes the natural logarithm. (See Math Refresher A for a review of logarithms.) In
particular, if Du 5 0, then
Notice how we multiply b1 by 100 to get the percentage change in wage given one additional
year of education. Because the percentage change in wage is the same for each additional year of
education, the change in wage for an extra year of education increases as education increases; in
other words, (2.42) implies an increasing return to education. By exponentiating (2.42), we can write
wage 5 exp 1 b0 1 b1educ 1 u 2 . This equation is graphed in Figure 2.6, with u 5 0.
58860_ch02_hr_019-065.indd 37
10/18/18 4:05 PM
38
PART 1 Regression Analysis with Cross-Sectional Data
Figure 2.6 wage 5 exp 1 b0 1 b1educ 2 , with b1 . 0.
wage
0
Example 2.10
educ
A Log Wage Equation
Using the same data as in Example 2.4, but using log(wage) as the dependent variable, we obtain the
following relationship:
log 1 wage 2 5 0.584 1 0.083 educ
[2.44]
n 5 526, R2 5 0.186.
The coefficient on educ has a percentage interpretation when it is multiplied by 100: wage increases
by 8.3% for every additional year of education. This is what economists mean when they refer to the
“return to another year of education.”
It is important to remember that the main reason for using the log of wage in (2.42) is to impose a
constant percentage effect of education on wage. Once equation (2.44) is obtained, the natural log of
wage is rarely mentioned. In particular, it is not correct to say that another year of education increases
log(wage) by 8.3%.
The intercept in (2.44) is not very meaningful, because it gives the predicted log(wage), when
educ 5 0. The R-squared shows that educ explains about 18.6% of the variation in log(wage) (not
wage). Finally, equation (2.44) might not capture all of the nonlinearity in the relationship between
wage and schooling. If there are “diploma effects,” then the twelfth year of education—graduation
from high school—could be worth much more than the eleventh year. We will learn how to allow for
this kind of nonlinearity in Chapter 7.
Estimating a model such as (2.42) is straightforward when using simple regression. Just define the
dependent variable, y, to be y 5 log 1 wage 2 . The independent variable is represented by x 5 educ. The
mechanics of OLS are the same as before: the intercept and slope estimates are given by the formulas
(2.17) and (2.19). In other words, we obtain b^ 0 and b^ 1 from the OLS regression of log(wage) on educ.
Another important use of the natural log is in obtaining a constant elasticity model.
58860_ch02_hr_019-065.indd 38
10/18/18 4:05 PM
CHAPTER 2 The Simple Regression Model
Example 2.11
39
CEO Salary and Firm Sales
We can estimate a constant elasticity model relating CEO salary to firm sales. The data set is the same
one used in Example 2.3, except we now relate salary to sales. Let sales be annual firm sales, measured in millions of dollars. A constant elasticity model is
log 1 salary 2 5 b0 1 b1log 1 sales 2 1 u,
[2.45]
log 1 salary 2 5 4.822 1 0.257 log 1 sales 2
[2.46]
where b1 is the elasticity of salary with respect to sales. This model falls under the simple regression
model by defining the dependent variable to be y = log(salary) and the independent variable to be
x 5 log 1 sales 2 . Estimating this equation by OLS gives
n 5 209, R2 5 0.211.
The coefficient of log(sales) is the estimated elasticity of salary with respect to sales. It implies that
a 1% increase in firm sales increases CEO salary by about 0.257%—the usual interpretation of an
elasticity.
The two functional forms covered in this section will often arise in the remainder of this text. We
have covered models containing natural logarithms here because they appear so frequently in applied
work. The interpretation of such models will not be much different in the multiple regression case.
It is also useful to note what happens to the intercept and slope estimates if we change
the units of measurement of the dependent variable when it appears in logarithmic form.
Because the change to logarithmic form approximates a proportionate change, it makes sense
that nothing happens to the slope. We can see this by writing the rescaled variable as c1yi for each
observation i. The original equation is log 1 yi 2 5 b0 1 b1xi 1 ui. If we add log 1 c1 2 to both sides,
we get log 1 c1 2 1 log 1 yi 2 5 3 log 1 c1 2 1 b0 4 1 b1xi 1 ui, or log 1 c1yi 2 5 3 log 1 c1 2 1 b0 4 1 b1xi 1 ui.
(Remember that the sum of the logs is equal to the log of their product, as shown in Math Refresher A.)
Therefore, the slope is still b1, but the intercept is now log 1 c1 2 1 b0. Similarly, if the independent
variable is log 1 x 2 , and we change the units of measurement of x before taking the log, the slope
remains the same, but the intercept changes. You will be asked to verify these claims in Problem 9.
We end this subsection by summarizing four combinations of functional forms available from
using either the original variable or its natural log. In Table 2.3, x and y stand for the variables in their
original form. The model with y as the dependent variable and x as the independent variable is called
the level-level model because each variable appears in its level form. The model with log 1 y 2 as the
dependent variable and x as the independent variable is called the log-level model. We will not explicitly discuss the level-log model here, because it arises less often in practice. In any case, we will see
examples of this model in later chapters.
The last column in Table 2.3 gives the interpretation of b1. In the log-level model, 100 ? b1 is
sometimes called the semi-elasticity of y with respect to x. As we mentioned in Example 2.11, in the
log-log model, b1 is the elasticity of y with respect to x. Table 2.3 warrants careful study, as we will
refer to it often in the remainder of the text.
Table 2.3 Summary of Functional Forms Involving Logarithms
Model
58860_ch02_hr_019-065.indd 39
Dependent Variable
Independent Variable
Interpretation of b1
Level-level
y
x
Dy 5 b1Dx
Level-log
y
log(x)
Log-level
log( y )
x
Dy 5 1 b1 /100 2 %Dx
Log-log
log( y )
log(x)
%Dy 5 b1%Dx
%Dy 5 1 100b1 2 Dx
10/18/18 4:05 PM
40
PART 1 Regression Analysis with Cross-Sectional Data
2-4c The Meaning of “Linear” Regression
The simple regression model that we have studied in this chapter is also called the simple
linear regression model. Yet, as we have just seen, the general model also allows for certain
nonlinear relationships. So what does “linear” mean here? You can see by looking at equation (2.1)
that y 5 b0 1 b1x 1 u. The key is that this equation is linear in the parameters b0 and b1. There are
no restrictions on how y and x relate to the original explained and explanatory variables of interest.
As we saw in Examples 2.10 and 2.11, y and x can be natural logs of variables, and this is quite common in applications. But we need not stop there. For example, nothing prevents us from using simple
regression to estimate a model such as cons 5 b0 1 b1 !inc 1 u, where cons is annual consumption
and inc is annual income.
Whereas the mechanics of simple regression do not depend on how y and x are defined, the
interpretation of the coefficients does depend on their definitions. For successful empirical work, it
is much more important to become proficient at interpreting coefficients than to become efficient at
computing formulas such as (2.19). We will get much more practice with interpreting the estimates in
OLS regression lines when we study multiple regression.
Plenty of models cannot be cast as a linear regression model because they are not linear in their
parameters; an example is cons 5 1/ 1 b0 1 b1inc 2 1 u. Estimation of such models takes us into the
realm of the nonlinear regression model, which is beyond the scope of this text. For most applications, choosing a model that can be put into the linear regression framework is sufficient.
2-5 Expected Values and Variances of the OLS Estimators
In Section 2-1, we defined the population model y 5 b0 1 b1x 1 u, and we claimed that the key
assumption for simple regression analysis to be useful is that the expected value of u given any value
of x is zero. In Sections 2-2, 2-3, and 2-4, we discussed the algebraic properties of OLS estimation.
We now return to the population model and study the statistical properties of OLS. In other words, we
now view b^ 0 and b^ 1 as estimators for the parameters b0 and b1 that appear in the population model.
This means that we will study properties of the distributions of b^ 0 and b^ 1 over different random
samples from the population. (Math Refresher C contains definitions of estimators and reviews some
of their important properties.)
2-5a Unbiasedness of OLS
We begin by establishing the unbiasedness of OLS under a simple set of assumptions. For future reference, it is useful to number these assumptions using the prefix “SLR” for simple linear regression.
The first assumption defines the population model.
Assumption SLR.1Linear in Parameters
In the population model, the dependent variable, y, is related to the independent variable, x, and the
error (or disturbance), u, as
y 5 b0 1 b1x 1 u,[2.47]
where b0 and b1 are the population intercept and slope parameters, respectively.
To be realistic, y, x, and u are all viewed as random variables in stating the population model. We discussed the interpretation of this model at some length in Section 2-1 and gave several examples. In the
previous section, we learned that equation (2.47) is not as restrictive as it initially seems; by choosing
58860_ch02_hr_019-065.indd 40
10/18/18 4:05 PM
CHAPTER 2 The Simple Regression Model
41
y and x appropriately, we can obtain interesting nonlinear relationships (such as constant elasticity
models).
We are interested in using data on y and x to estimate the parameters b0 and, especially, b1. We
assume that our data were obtained as a random sample. (See Math Refresher C for a review of random sampling.)
Assumption SLR.2
Random Sampling
We have a random sample of size n, 5 1 xi, yi 2 : i 5 1, 2, c, n 6 , following the population model in
equation (2.47).
We will have to address failure of the random sampling assumption in later chapters that deal with
time series analysis and sample selection problems. Not all cross-sectional samples can be viewed as
outcomes of random samples, but many can be.
We can write (2.47) in terms of the random sample as
yi 5 b0 1 b1xi 1 ui, i 5 1, 2, c, n,
[2.48]
where ui is the error or disturbance for observation i (for example, person i, firm i, city i, and so on).
Thus, ui contains the unobservables for observation i that affect yi. The ui should not be confused with
the residuals, u^ i, that we defined in Section 2-3. Later on, we will explore the relationship between the
errors and the residuals. For interpreting b0 and b1 in a particular application, (2.47) is most informative, but (2.48) is also needed for some of the statistical derivations.
The relationship (2.48) can be plotted for a particular outcome of data as shown in Figure 2.7.
As we already saw in Section 2-2, the OLS slope and intercept estimates are not defined unless
we have some sample variation in the explanatory variable. We now add variation in the xi to our list
of assumptions.
Figure 2.7 Graph of yi 5 b0 1 b1xi 1 ui.
y
yi
PRF
E(y x) 5 0 1
ui
x
1
u1
y1
x1
58860_ch02_hr_019-065.indd 41
xi
x
10/18/18 4:05 PM
42
PART 1 Regression Analysis with Cross-Sectional Data
Assumption SLR.3
Sample Variation in the Explanatory Variable
The sample outcomes on x, namely, 5 xi, i 5 1, c, n 6 , are not all the same value.
This is a very weak assumption—certainly not worth emphasizing, but needed nevertheless. If x
varies in the population, random samples on x will typically contain variation, unless the population
variation is minimal or the sample size is small. Simple inspection of summary statistics on xi reveals
whether Assumption SLR.3 fails: if the sample standard deviation of xi is zero, then Assumption
SLR.3 fails; otherwise, it holds.
Finally, in order to obtain unbiased estimators of b0 and b1, we need to impose the zero conditional mean assumption that we discussed in some detail in Section 2-1. We now explicitly add it to
our list of assumptions.
Assumption SLR.4
Zero Conditional Mean
The error u has an expected value of zero given any value of the explanatory variable. In other words,
E 1 u 0 x 2 5 0.
For a random sample, this assumption implies that E 1 ui 0 xi 2 5 0, for all i 5 1, 2, c, n.
In addition to restricting the relationship between u and x in the population, the zero conditional
mean assumption—coupled with the random sampling assumption—allows for a convenient technical
simplification. In particular, we can derive the statistical properties of the OLS estimators as conditional
on the values of the xi in our sample. Technically, in statistical derivations, conditioning on the sample
values of the independent variable is the same as treating the xi as fixed in repeated samples, which
we think of as follows. We first choose n sample values for x1, x2, c, xn. (These can be repeated.)
Given these values, we then obtain a sample on y (effectively by obtaining a random sample of the ui).
Next, another sample of y is obtained, using the same values for x1, x2, c, xn. Then another sample
of y is obtained, again using the same x1, x2, c, xn. And so on.
The fixed-in-repeated-samples scenario is not very realistic in nonexperimental contexts. For
instance, in sampling individuals for the wage-education example, it makes little sense to think
of choosing the values of educ ahead of time and then sampling individuals with those particular
levels of education. Random sampling, where individuals are chosen randomly and their wage and
education are both recorded, is representative of how most data sets are obtained for empirical
analysis in the social sciences. Once we assume that E 1 u 0 x 2 5 0, and we have random sampling,
nothing is lost in derivations by treating the xi as nonrandom. The danger is that the fixed-inrepeated-samples assumption always implies that ui and xi are independent. In deciding when simple regression analysis is going to produce unbiased estimators, it is critical to think in terms of
Assumption SLR.4.
Now, we are ready to show that the OLS estimators are unbiased. To this end, we use the fact
n
n
that g i51 1 xi 2 x 2 1 yi 2 y 2 5 g i51 1 xi 2 x 2 yi (see Math Refresher A) to write the OLS slope estimator in equation (2.19) as
a 1 xi 2 x 2 yi
n
b^ 1 5
i51
n
2
a 1 xi 2 x 2
.
[2.49]
i51
Because we are now interested in the behavior of b^ 1 across all possible samples, b^ 1 is properly viewed
as a random variable.
58860_ch02_hr_019-065.indd 42
10/18/18 4:05 PM
CHAPTER 2 The Simple Regression Model
43
We can write b^ 1 in terms of the population coefficient and errors by substituting the right-hand
side of (2.48) into (2.49). We have
a 1 xi 2 x 2 yi
n
b^ 1 5
a 1 xi 2 x 2 1 b0 1 b1xi 1 ui 2
n
i51
5
SSTx
i51
SSTx
,
[2.50]
n
where we have defined the total variation in xi as SSTx 5 g i51 1 xi 2 x 2 2 to simplify the notation.
(This is not quite the sample variance of the xi because we do not divide by n 2 1.) Using the algebra
of the summation operator, write the numerator of b^ 1 as
a 1 xi 2 x 2 b0 1 a 1 xi 2 x 2 b1xi 1 a 1 xi 2 x 2 ui
n
n
i51
i51
n
i51
5 b0 a 1 xi 2 x 2 1 b1 a 1 xi 2 x 2 xi 1 a 1 xi 2 x 2 ui.
n
n
n
[2.51]
n
n
n
As shown in Math Refresher A, g i51 1 xi 2 x 2 5 0 and g i51 1 xi 2 x 2 xi 5 g i51 1 xi 2 x 2 2 5 SSTx.
n
^
Therefore, we can write the numerator of b1 as b1SSTx 1 g i51 1 xi 2 x 2 ui. Putting this over the denominator
gives
i51
i51
a 1 xi 2 x 2 ui
i51
n
b^ 1 5 b1 1
i51
SSTx
5 b1 1 1 1/SSTx 2 a diui,
n
i51
[2.52]
where di 5 xi 2 x. We now see that the estimator b^ 1 equals the population slope, b1, plus a term that
is a linear combination in the errors 3 u1, u2, c, un 4 . Conditional on the values of xi, the randomness
in b^ 1 is due entirely to the errors in the sample. The fact that these errors are generally different from
zero is what causes b^ 1 to differ from b1.
Using the representation in (2.52), we can prove the first important statistical property of OLS.
Theorem
2.1
Unbiasedness of OLS:
Using Assumptions SLR.1 through SLR.4,
[2.53]
E 1 b^ 0 2 5 b0 and E 1 b^ 1 2 5 b1,
for any values of b0 and b1. In other words, b^ 0 is unbiased for b0, and b^ 1 is unbiased for b1.
PROOF: In this proof, the expected values are conditional on the sample values of the independent
­variable. Because SSTx and di are functions only of the xi, they are nonrandom in the conditioning.
Therefore, from (2.52), and keeping the conditioning on 5 x1, x2, c, xn 6 implicit, we have
E 1 b^ 1 2 5 b1 1 E 3 1 1/SSTx 2 a di ui 4 5 b1 1 1 1/SSTx 2 a E 1 di ui 2
n
n
i51
i51
5 b1 1 1 1/SSTx 2 a di E 1 Ui 2 5 b1 1 1 1/SSTx 2 a di ? 0 5 b1,
n
n
i51
i51
where we have used the fact that the expected value of each ui (conditional on 5 x1, x2, c, xn 6 2 is
zero under Assumptions SLR.2 and SLR.4. Because unbiasedness holds for any outcome on
5 x1, x2, c, xn 6 , unbiasedness also holds without conditioning on 5 x1, x2, c, xn 6 .
The proof for b^ 0 is now straightforward. Average (2.48) across i to get y 5 b0 1 b1x 1 u, and plug
this into the formula for b^ 0:
b^ 0 5 y 2 b^ 1x 5 b0 1 b1x 1 u 2 b^ 1x 5 b0 1 1 b1 2 b^ 1 2 x 1 u.
58860_ch02_hr_019-065.indd 43
10/18/18 4:05 PM
44
PART 1 Regression Analysis with Cross-Sectional Data
Then, conditional on the values of the xi,
E 1 b^ 0 2 5 b0 1 E 3 1 b1 2 b^ 1 2 x 4 1 E 1 u 2 5 b0 1 E 3 1 b1 2 b^ 1 2 4 x,
because E 1 u 2 5 0 by Assumptions SLR.2 and SLR.4. But, we showed that E 1 b^ 1 2 5 b1, which implies
that E 3 1 b^ 1 2 b1 2 4 5 0. Thus, E 1 b^ 0 2 5 b0. Both of these arguments are valid for any values of b0 and
b1, and so we have established unbiasedness.
Remember that unbiasedness is a feature of the sampling distributions of b^ 1 and b^ 0, which says nothing about the estimate that we obtain for a given sample. We hope that, if the sample we obtain is somehow
“typical,” then our estimate should be “near” the population value. Unfortunately, it is always possible that
we could obtain an unlucky sample that would give us a point estimate far from b1, and we can never
know for sure whether this is the case. You may want to review the material on unbiased estimators in Math
Refresher C, especially the simulation exercise in Table C.1 that illustrates the concept of unbiasedness.
Unbiasedness generally fails if any of our four assumptions fail. This means that it is important to
think about the veracity of each assumption for a particular application. Assumption SLR.1 requires
that y and x be linearly related, with an additive disturbance. This can certainly fail. But we also know
that y and x can be chosen to yield interesting nonlinear relationships. Dealing with the failure of
(2.47) requires more advanced methods that are beyond the scope of this text.
Later, we will have to relax Assumption SLR.2, the random sampling assumption, for time series
analysis. But what about using it for cross-sectional analysis? Random sampling can fail in a cross
section when samples are not representative of the underlying population; in fact, some data sets are
constructed by intentionally oversampling different parts of the population. We will discuss problems
of nonrandom sampling in Chapters 9 and 17.
As we have already discussed, Assumption SLR.3 almost always holds in interesting regression
applications. Without it, we cannot even obtain the OLS estimates.
The assumption we should concentrate on for now is SLR.4. If SLR.4 holds, the OLS estimators
are unbiased. Likewise, if SLR.4 fails, the OLS estimators generally will be biased. There are ways to
determine the likely direction and size of the bias, which we will study in Chapter 3.
The possibility that x is correlated with u is almost always a concern in simple regression analysis with nonexperimental data, as we indicated with several examples in Section 2-1. Using simple
regression when u contains factors affecting y that are also correlated with x can result in spurious
correlation: that is, we find a relationship between y and x that is really due to other unobserved factors that affect y and also happen to be correlated with x.
Example 2.12
Student Math Performance and the School Lunch Program
Let math10 denote the percentage of tenth graders at a high school receiving a passing score on a standardized mathematics exam. Suppose we wish to estimate the effect of the federally funded school
lunch program on student performance. If anything, we expect the lunch program to have a positive
ceteris paribus effect on performance: all other factors being equal, if a student who is too poor to eat
regular meals becomes eligible for the school lunch program, his or her performance should improve.
Let lnchprg denote the percentage of students who are eligible for the lunch program. Then, a simple
regression model is
math10 5 b0 1 b1 lnchprg 1 u,
[2.54]
where u contains school and student characteristics that affect overall school performance. Using the
data in MEAP93 on 408 Michigan high schools for the 1992–1993 school year, we obtain
math10 5 32.14 2 0.319 lnchprg
n 5 408, R2 5 0.171.
58860_ch02_hr_019-065.indd 44
10/18/18 4:05 PM
CHAPTER 2 The Simple Regression Model
45
This equation predicts that if student eligibility in the lunch program increases by 10 percentage
points, the percentage of students passing the math exam falls by about 3.2 percentage points. Do
we really believe that higher participation in the lunch program actually causes worse performance?
Almost certainly not. A better explanation is that the error term u in equation (2.54) is correlated with
lnchprg. In fact, u contains factors such as the poverty rate of children attending school, which affects
student performance and is highly correlated with eligibility in the lunch program. Variables such as
school quality and resources are also contained in u, and these are likely correlated with lnchprg. It
is important to remember that the estimate –0.319 is only for this particular sample, but its sign and
magnitude make us suspect that u and x are correlated, so that simple regression is biased.
In addition to omitted variables, there are other reasons for x to be correlated with u in the simple
regression model. Because the same issues arise in multiple regression analysis, we will postpone a
systematic treatment of the problem until then.
2-5b Variances of the OLS Estimators
In addition to knowing that the sampling distribution of b^ 1 is centered about b1 (b^ 1 is unbiased), it is
important to know how far we can expect b^ 1 to be away from b1 on average. Among other things, this
allows us to choose the best estimator among all, or at least a broad class of, unbiased estimators. The
measure of spread in the distribution of b^ 1 (and b^ 0) that is easiest to work with is the variance or its
square root, the standard deviation. (See Math Refresher C for a more detailed discussion.)
It turns out that the variance of the OLS estimators can be computed under Assumptions SLR.1
through SLR.4. However, these expressions would be somewhat complicated. Instead, we add an
assumption that is traditional for cross-sectional analysis. This assumption states that the variance of
the unobservable, u, conditional on x, is constant. This is known as the homoskedasticity or “constant variance” assumption.
Assumption SLR.5
Homoskedasticity
The error u has the same variance given any value of the explanatory variable. In other words,
Var 1 u 0 x 2 5 s2.
We must emphasize that the homoskedasticity assumption is quite distinct from the zero conditional mean assumption, E 1 u 0 x 2 5 0. Assumption SLR.4 involves the expected value of u, while
Assumption SLR.5 concerns the variance of u (both conditional on x). Recall that we established the
unbiasedness of OLS without Assumption SLR.5: the homoskedasticity assumption plays no role in
showing that b^ 0 and b^ 1 are unbiased. We add Assumption SLR.5 because it simplifies the variance
calculations for b^ 0 and b^ 1 and because it implies that ordinary least squares has certain efficiency
properties, which we will see in Chapter 3. If we were to assume that u and x are independent, then
the distribution of u given x does not depend on x, and so E 1 u 0 x 2 5 E 1 u 2 5 0 and Var 1 u 0 x 2 5 s2.
But independence is sometimes too strong of an assumption.
Because Var 1 u 0 x 2 5 E 1 u2 0 x 2 2 3 E 1 u 0 x 2 4 2 and E 1 u 0 x 2 5 0, s2 5 E 1 u2 0 x 2 , which means s2 is
also the unconditional expectation of u2. Therefore, s2 5 E 1 u2 2 5 Var 1 u 2 , because E 1 u 2 5 0. In
other words, s2 is the unconditional variance of u, and so s2 is often called the error variance or
disturbance variance. The square root of s2, s, is the standard deviation of the error. A larger s means
that the distribution of the unobservables affecting y is more spread out.
58860_ch02_hr_019-065.indd 45
10/18/18 4:05 PM
46
PART 1 Regression Analysis with Cross-Sectional Data
Figure 2.8 The simple regression model under homoskedasticity.
f(y x)
y
E(y x) 5
x1
x2
0
1
x
1
x3
x
It is often useful to write Assumptions SLR.4 and SLR.5 in terms of the conditional mean and
conditional variance of y:
E 1 y 0 x 2 5 b0 1 b1x.
Var 1 y 0 x 2 5 s2.
[2.55]
[2.56]
In other words, the conditional expectation of y given x is linear in x, but the variance of y given x is
constant. This situation is graphed in Figure 2.8 where b0 . 0 and b1 . 0.
When Var 1 u 0 x 2 depends on x, the error term is said to exhibit heteroskedasticity (or nonconstant
variance). Because Var 1 u 0 x 2 5 Var 1 y 0 x 2 , heteroskedasticity is present whenever Var 1 y 0 x 2 is a function of x.
Example 2.13
Heteroskedasticity in a Wage Equation
In order to get an unbiased estimator of the ceteris paribus effect of educ on wage, we must assume that
E 1 u 0 educ 2 5 0, and this implies E 1 wage 0 educ 2 5 b0 1 b1educ. If we also make the homoskedasticity assumption, then Var 1 u 0 educ 2 5 s2 does not depend on the level of education, which is the same
as assuming Var 1 wage 0 educ 2 5 s2. Thus, while average wage is allowed to increase with education
level—it is this rate of increase that we are interested in estimating—the variability in wage about its
mean is assumed to be constant across all education levels. This may not be realistic. It is likely that
people with more education have a wider variety of interests and job opportunities, which could lead
to more wage variability at higher levels of education. People with very low levels of education have
fewer opportunities and often must work at the minimum wage; this serves to reduce wage variability
at low education levels. This situation is shown in Figure 2.9. Ultimately, whether Assumption SLR.5
holds is an empirical issue, and in Chapter 8 we will show how to test Assumption SLR.5.
58860_ch02_hr_019-065.indd 46
10/18/18 4:05 PM
CHAPTER 2 The Simple Regression Model
47
Figure 2.9 Var 1 wage 0 educ 2 increasing with educ.
f(wage educ)
wage
8
E(wage educ) 5
0 1
1educ
12
16
educ
With the homoskedasticity assumption in place, we are ready to prove the following:
Theorem
2.2
Sampling Variances of the OLS Estimators
Under Assumptions SLR.1 through SLR.5,
Var 1 b^ 1 2 5
s2
2
a 1 xi 2 x 2
n
5 s2 /SSTxr,
[2.57]
i51
and
s2n 21 a x2i
n
Var 1 b^ 0 2 5
i51
2
a 1 xi 2 x 2
n
,
[2.58]
i51
where these are conditional on the sample values 5 x1, c, xn 6 .
PROOF: We derive the formula for Var 1 b^ 1 2 , leaving the other derivation as Problem 10. The starting point
is equation 1 2.52 2 : b^ 1 5 b1 1 1 1/SSTx 2 g ni51di ui. Because b1 is just a constant, and we are conditioning on the xi, SSTx and di 5 xi 2 x are also nonrandom. Furthermore, because the ui are independent
58860_ch02_hr_019-065.indd 47
10/18/18 4:05 PM
48
PART 1 Regression Analysis with Cross-Sectional Data
random variables across i (by random sampling), the variance of the sum is the sum of the variances.
Using these facts, we have
Var 1 b^ 1 2 5 1 1/SSTx 2 2 Vara a di ui b 5 1 1/SSTx 2 2 a a d2i Var 1 ui 2 b
n
n
i51
i51
5 1 1/SSTx 2 2 a a di2s2 b
n
i51
3 because Var 1 ui 2 5 s2 for all i 4
5 s2 1 1/SSTx 2 2 a a di2 b 5 s2 1 1/SSTx 2 2SSTx 5 s2/SSTx,
n
i51
which is what we wanted to show.
Equations (2.57) and (2.58) are the “standard” formulas for simple regression analysis, which
are invalid in the presence of heteroskedasticity. This will be important when we turn to confidence
intervals and hypothesis testing in multiple regression analysis.
For most purposes, we are interested in Var 1 b^ 1 2 . It is easy to summarize how this variance
depends on the error variance, s2, and the total variation in 5 x1, x2, c, xn 6 , SSTx. First, the larger the
error variance, the larger is Var 1 b^ 1 2 . This makes sense because more variation in the unobservables
affecting y makes it more difficult to precisely estimate b1. On the other hand, more variability in the
independent variable is preferred: as the variability in the xi increases, the variance of b^ 1 decreases.
This also makes intuitive sense because the more spread out is the sample of independent variables,
the easier it is to trace out the relationship between E 1 y 0 x 2 and x; that is, the easier it is to estimate b1.
If there is little variation in the xi, then it can be hard to pinpoint how E 1 y 0 x 2 varies with x. As the
sample size increases, so does the total variation in the xi. Therefore, a larger sample size results in a
smaller variance for b^ 1.
This analysis shows that, if we are interested in
b
and
we have a choice, then we should choose the
G o i n g F u rt h e r 2 . 5
1
xi to be as spread out as possible. This is sometimes
Show that, when estimating b0, it is best to
possible with experimental data, but rarely do we
have x 5 0. What is Var 1 b^ 0 2 in this case?
have this luxury in the social sciences: usually, we
n
2
[Hint: For any sample of numbers, g i51 xi $
n
2
must take the xi that we obtain via random sampling.
g i51 1 xi 2 x 2 , with equality only if x 5 0. 4
Sometimes, we have an opportunity to obtain larger
sample sizes, although this can be costly.
For the purposes of constructing confidence intervals and deriving test statistics, we will need to
work with the standard deviations of b^ 1 and b^ 0, sd(b^ 1) and sd(b^ 0). Recall that these are obtained by
taking the square roots of the variances in (2.57) and (2.58). In particular, sd 1 b^ 1 2 5 s/"SSTx, where
s is the square root of s2, and "SSTx is the square root of SSTx.
2-5c Estimating the Error Variance
The formulas in (2.57) and (2.58) allow us to isolate the factors that contribute to Var 1 b^ 1 2 and Var 1 b^ 0 2 .
But these formulas are unknown, except in the extremely rare case that s2 is known. Nevertheless, we
can use the data to estimate s2, which then allows us to estimate Var 1 b^ 1 2 and Var 1 b^ 0 2 .
This is a good place to emphasize the difference between the errors (or disturbances) and the
residuals, as this distinction is crucial for constructing an estimator of s2. Equation (2.48) shows how
to write the population model in terms of a randomly sampled observation as yi 5 b0 1 b1xi 1 ui,
where ui is the error for observation i. We can also express yi in terms of its fitted value and residual as
in equation 1 2.32 2 : yi 5 b^ 0 1 b^ 1xi 1 u^ i. Comparing these two equations, we see that the error shows
58860_ch02_hr_019-065.indd 48
10/18/18 4:06 PM
CHAPTER 2 The Simple Regression Model
49
up in the equation containing the population parameters, b0 and b1. On the other hand, the residuals
show up in the estimated equation with b^ 0 and b^ 1. The errors are never observed, while the residuals
are computed from the data.
We can use equations (2.32) and (2.48) to write the residuals as a function of the errors:
or
u^ i 5 yi 2 b^ 0 2 b^ 1xi 5 1 b0 1 b1xi 1 ui 2 2 b^ 0 2 b^ 1xi,
u^ i 5 ui 2 1 b^ 0 2 b0 2 2 1 b^ 1 2 b1 2 xi.
[2.59]
Although the expected value of b^ 0 equals b0, and similarly for b^ 1, u^ i is not the same as ui. The difference between them does have an expected value of zero.
Now that we understand the difference between the errors and the residuals, we can return to estin
mating s2. First, s2 5 E 1 u2 2 , so an unbiased “estimator” of s2 is n21 g i51 u2i . Unfortunately, this is not
a true estimator, because we do not observe the errors ui. But, we do have estimates of the ui, namely,
n
the OLS residuals u^ i. If we replace the errors with the OLS residuals, we have n21 g i51 u^ 2i 5 SSR/n.
This is a true estimator, because it gives a computable rule for any sample of data on x and y. One
slight drawback to this estimator is that it turns out to be biased (although for large n the bias is
small). Because it is easy to compute an unbiased estimator, we use that instead.
The estimator SSR/n is biased essentially because it does not account for two restrictions that
must be satisfied by the OLS residuals. These restrictions are given by the two OLS first order
conditions:
a u^ i 5 0,
n
i51
a xiu^ i 5 0.
n
i51
[2.60]
One way to view these restrictions is this: if we know n 2 2 of the residuals, we can always get the
other two residuals by using the restrictions implied by the first order conditions in (2.60). Thus,
there are only n 2 2 degrees of freedom in the OLS residuals, as opposed to n degrees of freedom
in the errors. It is important to understand that if we replace u^ i with ui in (2.60), the restrictions would
no longer hold.
The unbiased estimator of s2 that we will use makes a degrees of freedom adjustment:
s^ 2 5
n
1
2
a u^ i 5 SSR/ 1 n 2 2 2 .
1 n 2 2 2 i51
[2.61]
(This estimator is sometimes denoted as S 2, but we continue to use the convention of putting “hats”
over estimators.)
Theorem
2.3
Unbiased Estimation of s2
Under Assumptions SLR.1 through SLR.5,
E 1 s^ 2 2 5 s2.
PROOF: If we average equation (2.59) across all i and use the fact that the OLS residuals average out to zero,
we have 0 5 u 2 1 b^ 0 2 b0 2 2 1 b^ 1 2 b1 2 x; subtracting this from (2.59) gives u^ i 5 1 ui 2 u 2 2 1 b^ 1 2 b1 2
1 xi 2 x 2 . Therefore, u^ 2i 5 1 ui 2 u 2 2 1 1 b^ 1 2 b1 2 2 1 xi 2 x 2 2 2 2 1 ui 2 u 2 1 b^ 1 2 b1 2 1 xi 2 x 2 . Summing
n
n
across all i gives g i51u^ 2i 5 g ni51 1 ui 2 u 2 2 1 1 b^ 1 2 b1 2 2 g ni51 1 xi 2 x 2 2 2 2 1 b^ 1 2 b1 2 g i51 ui 1 xi 2 x 2 .
Now, the expected value of the first term is 1 n 2 1 2 s2, something that is shown in Math Refresher C. The
expected value of the second term is simply s2 because E 3 1 b^ 1 2 b1 2 2 4 5 Var 1 b^ 1 2 5 s2/SSTx. Finally,
the third term can be written as 22 1 b^ 1 2 b1 2 2SSTx; taking expectations gives 22s2. Putting these three
n
terms ­together gives E 1 g i51u^ 2i 2 5 1 n 2 1 2 s2 1 s2 2 2s2 5 1 n 2 2 2 s2, so that E 3 SSR/ 1 n 2 2 2 4 5 s2.
58860_ch02_hr_019-065.indd 49
10/18/18 4:06 PM
50
PART 1 Regression Analysis with Cross-Sectional Data
If s^ 2 is plugged into the variance formulas (2.57) and (2.58), then we have unbiased estimators of
Var 1 b^ 1 2 and Var 1 b^ 0 2 . Later on, we will need estimators of the standard deviations of b^ 1 and b^ 0, and
this requires estimating s. The natural estimator of s is
s^ 5 "s^ 2
[2.62]
and is called the standard error of the regression (SER). (Other names for s^ are the standard error
of the estimate and the root mean squared error, but we will not use these.) Although s^ is not an unbiased estimator of s, we can show that it is a consistent estimator of s (see Math Refresher C), and it
will serve our purposes well.
The estimate s^ is interesting because it is an estimate of the standard deviation in the unobservables
affecting y; equivalently, it estimates the standard deviation in y after the effect of x has been taken out.
Most regression packages report the value of s^ along with the R-squared, intercept, slope, and other OLS
statistics (under one of the several names listed above). For now, our primary interest is in using s^ to estimate the standard deviations of b^ 0 and b^ 1. Because sd 1 b^ 1 2 5 s/"SSTx, the natural estimator of sd 1 b^ 1 2 is
se 1 b^ 1 2 5 s^ /"SSTx 5 s^ / a a 1 xi 2 x 2 2 b ;
n
1/2
i51
this is called the standard error of b^ 1. Note that se 1 b^ 1 2 is viewed as a random variable when we think of
running OLS over different samples of y; this is true because s^ varies with different samples. For a given
sample, se 1 b^ 1 2 is a number, just as b^ 1 is simply a number when we compute it from the given data.
Similarly, se 1 b^ 0 2 is obtained from sd 1 b^ 0 2 by replacing s with s^ . The standard error of any estimate gives us an idea of how precise the estimator is. Standard errors play a central role throughout
this text; we will use them to construct test statistics and confidence intervals for every econometric
procedure we cover, starting in Chapter 4.
2-6 Regression through the Origin and Regression on a Constant
In rare cases, we wish to impose the restriction that, when x 5 0, the expected value of y is zero.
There are certain relationships for which this is reasonable. For example, if income (x) is zero, then
income tax revenues (y) must also be zero. In addition, there are settings where a model that originally
has a nonzero intercept is transformed into a model without an intercept.
&
Formally, we now choose a slope estimator, which we call b1, and a line of the form
&
&
[2.63]
|
|
where the tildes over b1 and y are used to distinguish this problem from the much more common
problem of estimating an intercept along with a slope. Obtaining (2.63) is called regression through
&
the origin because the line (2.63) passes through the point x 5 0, y 5 0. To obtain the slope estimate
in (2.63), we still rely on the method of ordinary least squares, which in this case minimizes the sum
of squared residuals:
y 5 b1x,
2
a 1 yi 2 b1xi 2 .
n
i51
&
[2.64]
&
Using one-variable calculus, it can be shown that b1 must solve the first order condition:
a xi 1 yi 2 b1xi 2 5 0.
n
&
From this, we can solve for b1:
&
i51
a xiyi
[2.65]
n
&
b1 5
i51
n
a xi
2
,
[2.66]
i51
provided that not all the xi are zero, a case we rule out.
58860_ch02_hr_019-065.indd 50
10/18/18 4:06 PM
CHAPTER 2 The Simple Regression Model
51
|
Note how b1 compares with the slope estimate when we also estimate the intercept (rather than
set it equal to zero). These two estimates are the same if, and only if, x 5 0. [See equation (2.49) for
b^ 1.] Obtaining an estimate of b1 using regression through the origin is not done very often in applied
|
work, and for good reason: if the intercept b0 2 0, then b1 is a biased estimator of b1. You will be
asked to prove this in Problem 8.
In cases where regression through the origin is deemed appropriate, one must be careful in interpreting the R-squared that is typically reported with such regressions. Usually, unless stated otherwise,
the R-squared is obtained without removing the sample average of 5 yi: i 5 1, c, n 6 in obtaining
SST. In other words, the R-squared is computed as
2
a 1 yi 2 b1xi 2
n
12
|
i51
a yi
n
2
.
[2.67]
i51
The numerator here makes sense because it is the sum of squared residuals, but the denominator
acts as if we know the average value of y in the population is zero. One reason this version of the
R-squared is used is that if we use the usual total sum of squares, that is, we compute R-squared as
2
a 1 yi 2 b1xi 2
n
12
i51
n
|
2
a 1 yi 2 y 2
,
[2.68]
i51
it can actually be negative. If expression (2.68) is negative then it means that using the sample average
y to predict yi provides a better fit than using xi in a regression through the origin. Therefore, (2.68) is
actually more attractive than equation (2.67) because equation (2.68) tells us whether using x is better
than ignoring x altogether.
This discussion about regression through the origin, and different ways to measure goodnessof-fit, prompts another question: what happens if we only regress on a constant? That is, what if we
set the slope to zero (which means we need not even have an x) and estimate an intercept only? The
answer is simple: the intercept is y. This fact is usually shown in basic statistics, where it is shown
that the constant that produces the smallest sum of squared deviations is always the sample average.
In this light, equation (2.68) can be seen as comparing regression on x through the origin with regression only on a constant.
2-7 Regression on a Binary Explanatory Variable
Our discussion so far has centered on the case where the explanatory variable, x, has quantitative
meanings. A few examples include years of schooling, return on equity for a firm, and the percentage of students at a school eligible for the federal free lunch program. We know how to interpret the
slope coefficient in each case. We also discussed interpretation of the slope coefficient when we use
the logarithmic transformations of the explained variable, the explanatory variable, or both.
Simple regression can also be applied to the case where x is a binary variable, often called a
dummy variable in the context of regression analysis. As the name “binary variable” suggests, x
takes on only two values, zero and one. These two values are used to put each unit in the population
into one of two groups represented by x 5 0 and x 5 1. For example, we can use a binary variable to
describe whether a worker participates in a job training program. In the spirit of giving our variables
descriptive names, we might use train to indicate participation: train 5 1 means a person participates;
train 5 0 means the person does not. Given a data set, we add an i subscript, as usual, so traini indicates job training status for a randomly drawn person i.
58860_ch02_hr_019-065.indd 51
10/18/18 4:06 PM
52
PART 1 Regression Analysis with Cross-Sectional Data
If we have a dependent or response variable, y, what does it mean to have a simple regression
equation when x is binary? Consider again the equation
y = b0 1 b1x 1 u
but where now x is a binary variable. If we impose the zero conditional mean assumption SLR.4 then
we obtain
E(y 0 x) 5 b0 1 b1x 1 E(u 0 x) 5 b0 1 b1x,
[2.69]
just as in equation (2.8). The only difference now is that x can take on only two values. By plugging
the values zero and one into (2.69), it is easily seen that
E(y 0 x 5 0) 5 b0
It follows immediately that
[2.70]
E(y 0 x 5 1) 5 b0 1 b1.
[2.71]
b1 5 E(y 0 x 5 1) 2 E(y 0 x 5 0).
[2.72]
In other words, b1 is the difference in the average value of y over the subpopulations with x 5 1 and
x 5 0. As with all simple regression analyses, this difference can be descriptive or, in a case discussed
in the next subsection, b1 can be a causal effect of an intervention or a program.
As an example, suppose that every worker in an hourly wage industry is put into one of two
racial categories: white (or Caucasian) and nonwhite. (Clearly this is a very crude way to categorize
race, but it has been used in some contexts.) Define the variable white 5 1 if a person is classified as
Caucasian and zero otherwise. Let wage denote hourly wage. Then
b1 5 E(wage 0 white 5 1) 2 E(wage 0 white 5 0)
is the difference in average hourly wages between white and nonwhite workers. Equivalently,
E(wage 0 white) 5 b0 1 b1white.
Notice that b1 always has the interpretation that it is the difference in average wages between whites
and nonwhites. However, it does not necessarily measure wage discrimination because there are many
legitimate reasons wages can differ, and some of those—such as education levels—could differ, on
average, by race.
The mechanics of OLS do not change just because x is binary. Let {(xi, yi) : i 5 1, . . . , n} be the
sample of size n. The OLS intercept and slope estimators are always given by (2.16) and (2.19), respectively. The residuals always have zero mean and are uncorrelated with the xi in the sample. The definition of R-squared is unchanged. And so on. Nevertheless, because xi is binary, the OLS estimates have
a simple, sensible interpretation. Let y0 be the average of the yi with xi 5 0 and y1 the average when
xi 5 1. Problem 2.13 asks you to show that
b^ 0 5 y0
[2.73]
b^ 0 5 y1 2 y0.
[2.74]
For example, in the wage/race example, if we run the regression
wagei on whitei, i 5 1, . . . , n
then b^ 0 5 wage0, the average hourly wage for nonwhites, and b^ 1 5 wage1 2 wage0, the difference
in average hourly wages between whites and nonwhites. Generally, equation (2.74) shows that the
“slope” in the regression is the difference in means, which is a standard estimator from basic statistics
when comparing two groups.
The statistical properties of OLS are also unchanged when x is binary. In fact, nowhere is this
ruled out in the statements of the assumptions. Assumption SLR.3 is satisfied provided we see some
58860_ch02_hr_019-065.indd 52
10/18/18 4:06 PM
CHAPTER 2 The Simple Regression Model
53
zeros and some ones for xi in our sample. For example, in the wage/race example, we need to observe
some whites and some nonwhites in order to obtain b^ 1.
As with any simple regression analysis, the main concern is the zero conditional mean assumption,
SLR.4. In many cases, this condition will fail because x is systematically related to other factors that
affect y, and those other factors are necessarily part of u. We alluded to this above in discussing differences
in average hourly wage by race: education and workforce experience are two variables that affect hourly
wage that could systematically differ by race. As another example, suppose we have data on SAT scores
for students who did and did not take at least one SAT preparation course. Then x is a binary variable, say,
course, and the outcome variable is the SAT score, sat. The decision to take the preparation course could
be systematically related to other factors that are predictive of SAT scores, such as family income and
parents’ education. A comparison of average SAT scores between the two groups is unlikely to uncover
the causal effect of the preparation course. The framework covered in the next subsection allows us to
determine the special circumstances under which simple regression can uncover a causal effect.
2-7a Counterfactual Outcomes, Causality, and Policy Analysis
Having introduced the notion of a binary explanatory variable, now is a good time to provide a formal
framework for studying counterfactual or potential outcomes, as touched on briefly in Chapter 1. We
are particularly interested in defining a causal effect or treatment effect.
In the simplest case, we are interested in evaluating an intervention or policy that has only two
states of the world: a unit is subjected to the intervention or not. In other words, those not subject
to the intervention or new policy act as a control group and those subject to the intervention as the
treatment group. Using the potential outcomes framework introduced in Chapter 1, for each unit i
in the population we assume there are outcomes in both states of the world, yi (0) and yi (1). We will
never observe any unit in both states of the world but we imagine each unit in both states. For example, in studying a job training program, a person does or does not participate. Then yi (0) is earnings if
person i does not participate and yi (1) is labor earnings if i does participate. These outcomes are well
defined before the program is even implemented.
The causal effect, somewhat more commonly called the treatment effect, of the intervention for
unit i is simply
tei 5 yi(1) 2 yi(0),
[2.75]
the difference between the two potential outcomes. There are a couple of noteworthy items about tei.
First, it is not observed for any unit i because it depends on both counterfactuals. Second, it can be
negative, zero, or positive. It could be that the causal effect is negative for some units and positive for
others.
We cannot hope to estimate tei for each unit i. Instead, the focus is typically on the average
treatment effect (ATE), also called the average causal effect (ACE). The ATE is simply the average of the treatment effects across the entire population. (Sometimes for emphasis the ATE is called
the population average treatment effect.) We can write the ATE parameter as
tate 5 E[tei] 5 E[yi(1) 2 yi(0)] 5 E[yi(1)] 2 E[yi(0)],
[2.76]
where the final expression uses linearity of the expected value. Sometimes, to emphasize the population nature of tate we write tate 5 E[y(1) 2 y(0)], where [y(0), y(1)] are the two random variables
representing the counterfactual outcomes in the population.
For each unit i let xi be the program participation status—a binary variable. Then the observed
outcome, yi, can be written as
yi 5 (1 2 xi)yi(0) 1 xiyi (1),
[2.77]
which is just shorthand for yi 5 yi (0) if xi 5 0 and yi 5 yi (1) if xi 5 1. This equation precisely
describes why, given a random sample from the population, we observe only one of yi(0) and yi(1).
58860_ch02_hr_019-065.indd 53
10/18/18 4:06 PM
54
PART 1 Regression Analysis with Cross-Sectional Data
To see how to estimate the average treatment effect, it is useful to rearrange (2.77):
yi 5 yi(0) 1 [yi(1) 2 yi(0)]xi
[2.78]
Now impose a simple (and, usually, unrealistic) constant treatment effect. Namely, for all i,
yi(1) 5 t 1 yi(0),
[2.79]
or t 5 yi(1) 2 yi(0). Plugging this into (2.78) gives
yi 5 yi(0) 1 txi.
Now write yi(0) 5 a0 1 ui(0) where, by definition, a0 5 E[yi(0)] and E[ui(0)] 5 0. Plugging this
in gives
yi 5 a0 1 txi 1 ui(0).
[2.80]
If we define b0 5 a0, b1 5 t, and ui 5 ui(0) then the equation becomes exactly as in equation (2.48):
yi 5 b0 1 b1xi 1 ui,
where b1 5 t is the treatment (or causal) effect.
We can easily determine that the simple regression estimator, which we now know is the difference in means estimator, is unbiased for the treatment effect, t. If xi is independent of ui(0) then
E[ui(0) 0 xi] 5 0,
so that SLR.4 holds. We have already shown that SLR.1 holds in our derivation of (2.80). As usual,
we assume random sampling (SLR.2), and SLR.3 holds provided we have some treated units and
some control units, a basic requirement. It is pretty clear we cannot learn anything about the effect of
the intervention if all sampled units are in the control group or all are in the treatment group.
The assumption that xi is independent of ui(0) is the same as xi is independent of yi(0). This
assumption can be guaranteed only under random assignment, whereby units are assigned to the
treatment and control groups using a randomization mechanism that ignores any features of the individual units. For example, in evaluating a job training program, random assignment occurs if a coin
is flipped to determine whether a worker is in the control group or treatment group. (The coin can be
biased in the sense that the probability of a head need not be 0.5.) Random assignment can be compromised if units do not comply with their assignment.
Random assignment is the hallmark of a randomized controlled trial (RCT), which has long been
considered the gold standard for determing whether medical interventions have causal effects. RCTs generate the kind of experimental data of the type discussed in Chapter 1. In recent years, RCTs have become
more popular in certain fields in economics, such as development economics and behavioral economics.
Unfortunately, RCTs can be very expensive to implement, and in many cases randomizing subjects into
control and treatment groups raises ethical issues. (For example, if giving low-income families access to
free health care improves child health outcomes then randomizing some families into the control group
means those children will have, on average, worse health outcomes than they could have otherwise.)
Even though RCTs are not always feasible for answering particular questions in economics and
other fields, it is a good idea to think about the experiment one would run if random assignment were
possible. Working through the simple thought experiment typically ensures that one is asking a sensible question before gathering nonexperimental data. For example, if we want to study the effects
of Internet access in rural areas on student performance, we might not have the resources (or ethical
clearance) to randomly assign Internet access to some students and not others. Nevertheless, thinking
about how such an experiment would be implemented sharpens our thinking about the potential outcomes framework and what we mean by the treatment effect.
Our discussion of random assignment so far shows that, in the context of a constant treatment
effect, the simple difference-in-means estimator, y1 2 y0, is unbiased for t. We can easily relax the
constant treatment effect assumption. In general, the individual treatment effect can be written as
tei 5 yi(1) 2 yi(0) 5 tate 1 [ui(1) 2 ui(0)],
58860_ch02_hr_019-065.indd 54
[2.81]
10/18/18 4:06 PM
CHAPTER 2 The Simple Regression Model
55
where yi(1) 5 a1 1 ui(1) and tate 5 a1 2 a0. It is helpful to think of tate as the average across the
entire population and ui(1 ) 2 ui(0) as the deviation from the population average for unit i. Plugging
(2.81) into (2.78) gives
yi 5 a0 1 tatexi 1 ui(0) 1 [ui(1) 2 ui(0)]xi ; a0 1 tatexi 1 ui,
[2.82]
where the error term is now
ui 5 ui(0) 1 [ui(1) 2 ui(0)]xi.
The random assignment assumption is now that xi is independent of [ui(0), ui(1)]. Even though ui
depends on xi, the zero conditional mean assumption holds:
E(ui 0 xi) 5 E[ui(0) 0 xi] 1 E[ui(1) 2 ui(0) 0 xi]xi
5 0 1 0 ? xi 5 0.
We have again verified SLR.4, and so we conclude that the simple OLS estimator is unbiased for a0 and
tate,where t^ ate is the difference-in-means estimator. [The error ui is not independent of xi. In particular,
as shown in Problem 2.17, Var(ui 0 xi) differs across xi 5 1 and xi 5 0 if the variances of the potential outcomes differ. But remember, Assumption SLR.5 is not used to show the OLS estimators are unbiased.]
The fact that the simple regression estimator produces an unbiased estimator tate when the treatment effects can vary arbitrarily across individual units is a very powerful result. However, it relies
heavily on random assignment. Starting in Chapter 3, we will see how multiple regression analysis
can be used when pure random assignment does not hold. Chapter 20, available as an online supplement, contains an accessible survey of advanced methods for estimating treatment effects.
EXAMPLE 2.14
Evaluating a Job Training Program
The data in JTRAIN2 are from an old, experimental job training program, where men with poor labor
market histories were assigned to control and treatment groups. This data set has been used widely in
the program evaluation literature to compare estimates from nonexperimental programs. The training
assignment indicator is train and here we are interested in the outcome re78, which is (real) earnings in
1978 measured in thousands of dollars. Of the 445 men in the sample, 185 participated in the program
in a period prior to 1978; the other 260 men comprise the control group.
The simple regression gives
re78 5 4.55 1 1.79 train
n 5 445, R2 5 0.018.
From the earlier discussion, we know that 1.79 is the difference in average re78 between the treated
and control groups, so men who participated in the program earned an average of $1,790 more than
the men who did not. This is an economically large effect, as the dollars are 1978 dollars. Plus, the
average earnings for men who did not participate is $4,550; in percentage terms, the gain in average
earnings is about 39.3%, which is large. (We would need to know the costs of the program to do a
­benefit-cost analysis, but the benefits are nontrivial.)
Remember that the fundamental issue in program evaluation is that we do not observe any of the
units in both states of the world. In this example, we only observe one of the two earnings outcomes
for each men. Neverthless, random assignment into the treatment and control groups allows us to get
an unbiased estimator of the average treatment effect.
Two final comments on this example. First, notice the very small R-squared: the training participation indicator explains less than two percent of the variation in re78 in the sample. We should
not be surprised: many other factors, including education, experience, intelligence, age, motivation,
and so on help determine labor market earnings. This is a good example to show how focusing on
R-squared is not only unproductive, but it can be harmful. Beginning students sometimes think a
small R-squared indicates “bias” in the OLS estimators. It does not. It simply means that the variance
in the unobservables, Var(u), is large relative to Var(y). In this example, we know that Assumptions
58860_ch02_hr_019-065.indd 55
10/18/18 4:06 PM
56
PART 1 Regression Analysis with Cross-Sectional Data
SLR.1 to SLR.4 hold because of random assignment. Rightfully, none of these assumptions mentions
how large R-squared must be; it is immaterial for the notion of unbiasedness.
A second comment is that, while the estimated economic effect of $1,790 is large, we do not
know whether this estimate is statistically significant. We will come to this topic in Chapter 4.
Before ending this chapter, it is important to head off possible confusion about two different ways
the word “random” has been used in this subsection. First, the notion of random sampling is the one
introduced in Assumption SLR.2 (and also discussed in Math Refresher C). Random sampling means
that the data we obtain are independent, identically distributed draws from the population distribution
represented by the random variables (x, y). It is important to understand that random sampling is a separate concept from random assignment, which means that xi is determined independently of the counterfactuals [yi(0), yi(1)]. In Example 2.14, we obtained a random sample from the relevant population, and
the assignment to treatment and control is randomized. But in other cases, random assignment will not
hold even though we have random sampling. For example, it is relatively easy to draw a random sample
from a large population of college-bound students and obtain outcomes on their SAT scores and whether
they participated in an SAT preparation course. That does not mean that participation in a course is independent of the counterfactual outcomes. If we wanted to ensure independence between participation and
the potential outcomes, we would randomly assign the students to take a course or not (and insist that
students adhere to their assignments). If instead we obtain retrospective data—that is, we simply record
whether a student has taken a preparation course—then the independence assumption underlying RA is
unlikely to hold. But this has nothing to do with whether we obtained a random sample of students from
the population. The general point is that Assumptions SLR.2 and SLR.4 are very different.
Summary
We have introduced the simple linear regression model in this chapter, and we have covered its basic properties. Given a random sample, the method of ordinary least squares is used to estimate the slope and
intercept parameters in the population model. We have demonstrated the algebra of the OLS regression
line, including computation of fitted values and residuals, and the obtaining of predicted changes in the
dependent variable for a given change in the independent variable. In Section 2-4, we discussed two issues
of practical importance: (1) the behavior of the OLS estimates when we change the units of measurement
of the dependent variable or the independent variable and (2) the use of the natural log to allow for constant
elasticity and constant semi-elasticity models.
In Section 2-5, we showed that, under the four Assumptions SLR.1 through SLR.4, the OLS estimators are unbiased. The key assumption is that the error term u has zero mean, or average, given any value
of the independent variable x. Unfortunately, there are reasons to think this is false in many social science
applications of simple regression, where the omitted factors in u are often correlated with x. When we add
the assumption that the variance of the error given x is constant, we get simple formulas for the sampling
variances of the OLS estimators. As we saw, the variance of the slope estimator b^ 1 increases as the error
variance increases, and it decreases when there is more sample variation in the independent variable. We
also derived an unbiased estimator for s2 5 Var 1 u 2 .
In Section 2-6, we briefly discussed regression through the origin, where the slope estimator is
obtained under the assumption that the intercept is zero. Sometimes, this is useful, but it appears infrequently in applied work.
In Section 2-7 we covered the important case where x is a binary variable, and showed that
the OLS “slope” estimate is simply b^ 0 5 y1 2 y0, the difference in the averages of yi between the
xi = 1 and xi = 0 subsamples. We also discussed how, in the context of causal inference, b^ 1 is an
­unbiased estimator of the average treatment effect under random assignment into the control and
treatment groups. In Chapter 3 and beyond, we will study the case where the intervention or treatment
is not randomized, but depends on observed and even unobserved factors.
Much work is left to be done. For example, we still do not know how to test hypotheses about the population parameters, b0 and b1. Thus, although we know that OLS is unbiased for the population parameters
58860_ch02_hr_019-065.indd 56
10/18/18 4:06 PM
CHAPTER 2 The Simple Regression Model
57
under Assumptions SLR.1 through SLR.4, we have no way of drawing inferences about the population.
Other topics, such as the efficiency of OLS relative to other possible procedures, have also been omitted.
The issues of confidence intervals, hypothesis testing, and efficiency are central to multiple regression
analysis as well. Because the way we construct confidence intervals and test statistics is very similar for multiple regression—and because simple regression is a special case of multiple regression—our time is better spent
moving on to multiple regression, which is much more widely applicable than simple regression. Our purpose
in Chapter 2 was to get you thinking about the issues that arise in econometric analysis in a fairly simple setting.
The Gauss-Markov Assumptions for Simple Regression
For convenience, we summarize the Gauss-Markov assumptions that we used in this chapter. It is important to remember that only SLR.1 through SLR.4 are needed to show b^ 0 and b^ 1 are unbiased. We added the
homoskedasticity assumption, SLR.5, to obtain the usual OLS variance formulas (2.57) and (2.58).
Assumption SLR.1 (Linear in Parameters)
In the population model, the dependent variable, y, is related to the independent variable, x, and the error
(or disturbance), u, as
y 5 b0 1 b1x 1 u,
where b0 and b1 are the population intercept and slope parameters, respectively.
Assumption SLR.2 (Random Sampling)
We have a random sample of size n, 5 1 xi,yi 2 : i 5 1, 2, c, n 6 , following the population model in
Assumption SLR.1.
Assumption SLR.3 (Sample Variation in the Explanatory Variable)
The sample outcomes on x, namely, 5 xi, i 5 1, c, n 6 , are not all the same value.
Assumption SLR.4 (Zero Conditional Mean)
The error u has an expected value of zero given any value of the explanatory variable. In other words,
Assumption SLR.5 (Homoskedasticity)
E 1 u 0 x 2 5 0.
The error u has the same variance given any value of the explanatory variable. In other words,
Var 1 u 0 x 2 5 s2.
Key Terms
Average Treatment Effect (ATE)
Average Causal Effect (ACE)
Binary (Dummy) Variable
Causal (Treatment) Effect
Coefficient of Determination
Constant Elasticity Model
Control Group
Control Variable
Covariate
Degrees of Freedom
Dependent Variable
Elasticity
Error Term (Disturbance)
Error Variance
Explained Sum of Squares (SSE)
Explained Variable
Explanatory Variable
58860_ch02_hr_019-065.indd 57
First Order Conditions
Fitted Value
Gauss-Markov Assumptions
Heteroskedasticity
Homoskedasticity
Independent Variable
Intercept Parameter
Mean Independent
OLS Regression Line
Ordinary Least Squares (OLS)
Population Regression Function (PRF)
Predicted Variable
Predictor Variable
Random Assignment
Randomized Controlled Trial (RCT)
Regressand
Regression through the Origin
Regressor
Residual
Residual Sum of Squares (SSR)
Response Variable
R-squared
Sample Regression Function (SRF)
Semi-elasticity
Simple Linear Regression Model
Slope Parameter
Standard Error of b^ 1
Standard Error of the Regression (SER)
Sum of Squared Residuals
Total Sum of Squares (SST)
Treatment Group
Zero Conditional Mean Assumption
10/18/18 4:06 PM
58
PART 1 Regression Analysis with Cross-Sectional Data
Problems
1 Let kids denote the number of children ever born to a woman, and let educ denote years of education
for the woman. A simple model relating fertility to years of education is
kids 5 b0 1 b1educ 1 u,
where u is the unobserved error.
(i)
(ii)
What kinds of factors are contained in u? Are these likely to be correlated with level of education?
Will a simple regression analysis uncover the ceteris paribus effect of education on fertility?
Explain.
2 In the simple linear regression model y 5 b0 1 b1x 1 u, suppose that E 1 u 2 2 0. Letting a0 5 E 1 u 2 ,
show that the model can always be rewritten with the same slope, but a new intercept and error, where
the new error has a zero expected value.
3 The following table contains the ACT scores and the GPA (grade point average) for eight college students. Grade point average is based on a four-point scale and has been rounded to one digit after the
decimal.
(i)
Student
GPA
ACT
1
2.8
21
2
3.4
24
3
3.0
26
4
3.5
27
5
3.6
29
6
3.0
25
7
2.7
25
8
3.7
30
Estimate the relationship between GPA and ACT using OLS; that is, obtain the intercept and
slope estimates in the equation
GPA 5 b^ 0 1 b^ 1ACT.
Comment on the direction of the relationship. Does the intercept have a useful interpretation
here? Explain. How much higher is the GPA predicted to be if the ACT score is increased by
five points?
(ii) Compute the fitted values and residuals for each observation, and verify that the residuals
(approximately) sum to zero.
(iii) What is the predicted value of GPA when ACT 5 20?
(iv) How much of the variation in GPA for these eight students is explained by ACT? Explain.
4 The data set BWGHT contains data on births to women in the United States. Two variables of interest
are the dependent variable, infant birth weight in ounces (bwght), and an explanatory variable, average
number of cigarettes the mother smoked per day during pregnancy (cigs). The following simple regression was estimated using data on n 5 1,388 births:
bwght 5 119.77 2 0.514 cigs
(i)
(ii)
58860_ch02_hr_019-065.indd 58
What is the predicted birth weight when cigs 5 0? What about when cigs 5 20 (one pack per
day)? Comment on the difference.
Does this simple regression necessarily capture a causal relationship between the child’s birth
weight and the mother’s smoking habits? Explain.
10/18/18 4:06 PM
CHAPTER 2 The Simple Regression Model
59
(iii) To predict a birth weight of 125 ounces, what would cigs have to be? Comment.
(iv) The proportion of women in the sample who do not smoke while pregnant is about .85. Does
this help reconcile your finding from part (iii)?
5 In the linear consumption function
cons 5 b^ 0 1 b^ 1inc,
the (estimated) marginal propensity to consume (MPC) out of income is simply the slope, b^ 1,
while the average propensity to consume (APC) is cons/inc 5 b^ 0 /inc 1 b^ 1. Using observations
for 100 families on annual income and consumption (both measured in dollars), the following
equation is obtained:
cons 5 2124.84 1 0.853 inc
n 5 100, R2 5 0.692.
(i) Interpret the intercept in this equation, and comment on its sign and magnitude.
(ii) What is the predicted consumption when family income is $30,000?
(iii) With inc on the x-axis, draw a graph of the estimated MPC and APC.
6 Using data from 1988 for houses sold in Andover, Massachusetts, from Kiel and McClain (1995),
the following equation relates housing price (price) to the distance from a recently built garbage incinerator (dist):
log 1 price 2 5 9.40 1 0.312 log 1 dist 2
n 5 135, R2 5 0.162.
(i)
(ii)
Interpret the coefficient on log(dist). Is the sign of this estimate what you expect it to be?
Do you think simple regression provides an unbiased estimator of the ceteris paribus
elasticity of price with respect to dist? (Think about the city’s decision on where to put
the incinerator.)
(iii) What other factors about a house affect its price? Might these be correlated with distance from
the incinerator?
7 Consider the savings function
sav 5 b0 1 b1inc 1 u, u 5 "inc . e,
where e is a random variable with E 1 e 2 5 0 and Var 1 e 2 5 s2e . Assume that e is independent
of inc.
Show that E 1 u 0 inc 2 5 0, so that the key zero conditional mean assumption (Assumption SLR.4)
is satisfied. [Hint: If e is independent of inc, then E 1 e 0 inc 2 5 E 1 e 2 .]
(ii) Show that Var 1 u 0 inc 2 5 s2e inc, so that the homoskedasticity Assumption SLR.5 is violated. In
particular, the variance of sav increases with inc. [Hint: Var 1 e 0 inc 2 5 Var 1 e 2 if e and inc are
independent.]
(iii) Provide a discussion that supports the assumption that the variance of savings increases with
family income.
(i)
8 Consider the standard simple regression model y 5 b0 1 b1x 1 u under the Gauss-Markov
Assumptions SLR.1 through SLR.5. The usual OLS estimators b^ 0 and b^ 1 are unbiased for their respec&
tive population parameters. Let b1 be the estimator of b1 obtained by assuming the intercept is zero
(see Section 2-6).
&
&
(i) Find E 1 b1 2 in terms of the xi, b0, and b1. Verify that b1 is unbiased for b1 when the population
&
intercept 1 b0 2 is zero. Are there other cases where b1 is unbiased?
&
(ii) Find the variance of b1. (Hint: The variance does not depend on b0.)
58860_ch02_hr_019-065.indd 59
10/18/18 4:06 PM
60
PART 1 Regression Analysis with Cross-Sectional Data
n
n
(iii) Show that Var 1 b1 2 # Var 1 b^ 1 2 . [Hint: For any sample of data, g i21x2i $ g i51 1 xi 2 x 2 2, with
strict inequality unless x 5 0.]
&
(iv) Comment on the tradeoff between bias and variance when choosing between b^ 1 and b1.
&
9 (i)Let b^ 0 and b^ 1 be the intercept and slope from the& regression of yi on xi, using n observations. Let
&
c1 and c2, with c2 2 0, be constants. Let b0 and b1 be the intercept and slope from the regression
&
&
of c1yi on c2xi. Show that b1 5 1 c1/c2 2 b^ 0 and b0 5 c1b^ 0, thereby verifying the claims on units of
&
measurement in Section 2-4. [Hint: To obtain b1, plug the scaled versions of x and y into (2.19).
&
Then, use (2.17)&for b0, being sure to plug in the scaled x and y and the correct slope.]
&
(ii) Now, let b0 and b1 be from the regression of 1 c1 1 yi 2 on 1 c2 1 xi 2 (with no restriction on c1 or c2).
&
&
Show that b1 5 b^ 1 and b0 5 b^ 0 1 c1 2 c2b^ 1.
(iii) Now, let b^ 0 and b^ 1 be the OLS estimates& from &the regression log 1 yi 2 on xi, where we must
assume yi . 0 for all i. For c1& . 0, let b0 and
b1 be the intercept and slope from the regression
&
of log 1 c1yi 2 on xi. Show that b1 5 b^ 1 and
b
5
log 1 c1 2 1 b^ 0.
0
&
&
(iv) Now, assuming that xi . 0 &for all i,& let b0 and b1 be the intercept and slope from the regression
of yi on log 1 c2xi 2 . How do b0 and b1 compare with the intercept and slope from the regression
of yi on log 1 xi 2 ?
10 Let b^ 0 and b^ 1 be the OLS intercept and slope estimators, respectively, and let u be the sample average
of the errors (not the residuals!).
n
(i) Show that b^ 1 can be written as b^ 1 5 b1 1 g i51wiui, where wi 5 di /SSTx and di 5 xi 2 x.
n
(ii) Use part (i), along with g i51wi 5 0, to show that b^ 1 and u are uncorrelated. [Hint: You are
­being asked to show that E 3 1 b^ 1 2 b1 2 # u 4 5 0. 4
(iii) Show that b^ 0 can be written as b^ 0 5 b0 1 u 2 1 b^ 1 2 b1 2 x.
(iv) Use parts (ii) and (iii) to show that Var 1 b^ 0 2 5 s2/n 1 s2 1 x 2 2/SSTx.
(v) Do the algebra to simplify the expression in part (iv) to equation (2.58).
n
[Hint: SSTx/n 5 n21 g i51x2i 2 1 x 2 2.]
11 Suppose you are interested in estimating the effect of hours spent in an SAT preparation course
(hours) on total SAT score (sat). The population is all college-bound high school seniors for a particular year.
(i) Suppose you are given a grant to run a controlled experiment. Explain how you would structure
the experiment in order to estimate the causal effect of hours on sat.
(ii) Consider the more realistic case where students choose how much time to spend in a preparation course, and you can only randomly sample sat and hours from the population. Write the
population model as
sat 5 b0 1 b1hours 1 u
where, as usual in a model with an intercept, we can assume E 1 u 2 5 0. List at least two factors
contained in u. Are these likely to have positive or negative correlation with hours?
(iii) In the equation from part (ii), what should be the sign of b1 if the preparation course is effective?
(iv) In the equation from part (ii), what is the interpretation of b0?
12 Consider the problem described at the end of Section 2-6, running a regression and only estimating an
intercept.
&
(i) Given a sample 5 yi: i 5 1, 2, c, n 6 , let b0 be the solution to
min a 1 yi 2 b0 2 2.
b
n
0
&
(ii)
58860_ch02_hr_019-065.indd 60
i51
Show that b0 5 y, that is, the sample average minimizes the sum of squared residuals. (Hint:
You may use one-variable calculus or you can show the result directly by adding and subtracting y inside the squared residual and then doing a little algebra.)
&
Define residuals ui 5 yi 2 y. Argue that these residuals always sum to zero.
10/18/18 4:06 PM
CHAPTER 2 The Simple Regression Model
61
13 Let y be any response variable and x a binary explanatory variable. Let {(xi, yi) : i 5 1, . . . , n} be a
sample of size n. Let n0 be the number of observations with xi 5 0 and n1 the number of observations
with xi 5 1. Let y0 be the average of the yi with xi 5 0 and y1 the average of the yi with xi 5 1.
(i) Explain why we can write
n0 5 a 1 1 2 xi 2 , n1 5 a xi.
n
n
i51
i51
Show that x 5 n1/n and (1 2 x) 5 n0/n. How do you interpret x ?
(ii) Argue that
21
y0 5 n21
0 a 1 1 2 xi 2 yi, y1 5 n1 a xi yi .
n
n
i51
i51
(iii) Show that the average of yi in the entire sample, y, can be written as a weighted average:
y 5 1 1 2 x 2 y0 1 x y1.
[Hint: Write yi 5 (1 2 xi)yi 1 xiyi.]
(iv) Show that when xi is binary,
n21 a x2i 2 1 x 2 2 5 x 1 1 2 x 2 .
n
i51
[Hint: When xi is binary, x2i 5 xi.]
(v) Show that
n21 a xi yi 2 x y 5 x 1 1 2 x 2 1 y1 2 y0 2 .
n
i51
(vi) Use parts (iv) and (v) to obtain (2.74).
(vii) Derive equation (2.73).
14 In the context of Problem 2.13, suppose yi is also binary. For concreteness, yi indicates whether
worker i is employed after a job training program, where yi 5 1 means has a job, yi 5 0 means does not
have a job. Here, xi indicates participation in the job training program. Argue that b^ 1 is the difference in
employment rates between those who participated in the program and those who did not.
15 Consider the potential outcomes framework from Section 2.7a, where yi(0) and yi(1) are the potential
outcomes in each treatment state.
(i) Show that if we could observe yi(0) and yi(1) for all i then an unbiased estimator of tate would be
n21 a 3 yi 1 1 2 2 yi 1 0 2 4 5 y 1 1 2 2 y 1 0 2 .
n
i51
This is sometimes called the sample average treatment effect.
(ii) Explain why the observed sample averages, y0 and y1, are not the same as y(0) and y(1),
­respectively, by writing y0 and y1 in terms of yi(0) and yi(1), respectively.
16 In the potential outcomes framework, suppose that program eligibility is randomly assigned but participation cannot be enforced. To formally describe this situation, for each person i, zi is the eligibility indicator and xi is the participation indicator. Randomized eligibility means zi is independent of
[yi(0), yi(1)] but xi might not satisfy the independence assumption.
(i) Explain why the difference in means estimator is generally no longer unbiased.
(ii) In the context of a job training program, what kind of individual behavior would cause bias?
17 In the potential outcomes framework with heterogeneous (nonconstant) treatment effect, write the
error as
ui 5 (1 2 xi)ui(0) 1 xiui(1).
Let s20 5 Var[ui(0)] and s21 5 Var[ui(1)]. Assume random assignment.
58860_ch02_hr_019-065.indd 61
10/18/18 4:06 PM
62
PART 1 Regression Analysis with Cross-Sectional Data
(i)
(ii)
Find Var(ui 0 xi).
When is Var(ui 0 xi) constant?
18 Let x be a binary explanatory variable and suppose P(x 5 1) 5 r for 0 , r , 1.
(i) If you draw a random sample of size n, find the probability—call it gn—that Assumption SLR.3
fails. [Hint: Find the probability of observing all zeros or all ones for the xi.] Argue that gn S 0
as n S `.
(ii) If r 5 0.5, compute the probablity in part (i) for n 5 10 and n 5 100. Discuss.
(iii) Do the calculations from part (ii) with r 5 0.9. How do your answers compare with part (ii)?
Computer Exercises
C1 The data in 401K are a subset of data analyzed by Papke (1995) to study the relationship between
participation in a 401(k) pension plan and the generosity of the plan. The variable prate is the percentage of eligible workers with an active account; this is the variable we would like to explain. The
measure of generosity is the plan match rate, mrate. This variable gives the average amount the firm
contributes to each worker’s plan for each $1 contribution by the worker. For example, if mrate 5 0.50,
then a $1 contribution by the worker is matched by a 50¢ contribution by the firm.
(i) Find the average participation rate and the average match rate in the sample of plans.
(ii) Now, estimate the simple regression equation
prate 5 b^ 0 1 b^ 1 mrate,
and report the results along with the sample size and R-squared.
(iii) Interpret the intercept in your equation. Interpret the coefficient on mrate.
(iv) Find the predicted prate when mrate 5 3.5. Is this a reasonable prediction? Explain what is
happening here.
(v) How much of the variation in prate is explained by mrate? Is this a lot in your opinion?
C2 The data set in CEOSAL2 contains information on chief executive officers for U.S. corporations. The
variable salary is annual compensation, in thousands of dollars, and ceoten is prior number of years as
company CEO.
(i) Find the average salary and the average tenure in the sample.
(ii) How many CEOs are in their first year as CEO (that is, ceoten 5 0)? What is the longest tenure
as a CEO?
(iii) Estimate the simple regression model
log 1 salary 2 5 b0 1 b1ceoten 1 u,
and report your results in the usual form. What is the (approximate) predicted percentage
­increase in salary given one more year as a CEO?
C3 Use the data in SLEEP75 from Biddle and Hamermesh (1990) to study whether there is a tradeoff
­between the time spent sleeping per week and the time spent in paid work. We could use either variable
as the dependent variable. For concreteness, estimate the model
sleep 5 b0 1 b1totwrk 1 u,
(i)
(ii)
58860_ch02_hr_019-065.indd 62
where sleep is minutes spent sleeping at night per week and totwrk is total minutes worked during the week.
Report your results in equation form along with the number of observations and R2. What does
the intercept in this equation mean?
If totwrk increases by 2 hours, by how much is sleep estimated to fall? Do you find this to be a
large effect?
10/18/18 4:06 PM
CHAPTER 2 The Simple Regression Model
63
C4 Use the data in WAGE2 to estimate a simple regression explaining monthly salary (wage) in terms of
IQ score (IQ).
(i) Find the average salary and average IQ in the sample. What is the sample standard deviation
of IQ? (IQ scores are standardized so that the average in the population is 100 with a standard
deviation equal to 15.)
(ii) Estimate a simple regression model where a one-point increase in IQ changes wage by a constant dollar amount. Use this model to find the predicted increase in wage for an increase in
IQ of 15 points. Does IQ explain most of the variation in wage?
(iii) Now, estimate a model where each one-point increase in IQ has the same percentage effect on
wage. If IQ increases by 15 points, what is the approximate percentage increase in predicted wage?
C5 For the population of firms in the chemical industry, let rd denote annual expenditures on research and
development, and let sales denote annual sales (both are in millions of dollars).
(i) Write down a model (not an estimated equation) that implies a constant elasticity between
rd and sales. Which parameter is the elasticity?
(ii) Now, estimate the model using the data in RDCHEM. Write out the estimated equation in the
usual form. What is the estimated elasticity of rd with respect to sales? Explain in words what
this elasticity means.
C6 We used the data in MEAP93 for Example 2.12. Now we want to explore the relationship between the
math pass rate (math10) and spending per student (expend).
(i) Do you think each additional dollar spent has the same effect on the pass rate, or does a diminishing effect seem more appropriate? Explain.
(ii) In the population model
math10 5 b0 1 b1 log 1 expend 2 1 u,
argue that b1/10 is the percentage point change in math10 given a 10% increase in expend.
(iii) Use the data in MEAP93 to estimate the model from part (ii). Report the estimated equation in
the usual way, including the sample size and R-squared.
(iv) How big is the estimated spending effect? Namely, if spending increases by 10%, what is the
estimated percentage point increase in math10?
(v) One might worry that regression analysis can produce fitted values for math10 that are greater
than 100. Why is this not much of a worry in this data set?
C7 Use the data in CHARITY [obtained from Franses and Paap (2001)] to answer the following questions:
(i) What is the average gift in the sample of 4,268 people (in Dutch guilders)? What percentage of
people gave no gift?
(ii) What is the average mailings per year? What are the minimum and maximum values?
(iii) Estimate the model
gift 5 b0 1 b1mailsyear 1 u
(iv)
(v)
by OLS and report the results in the usual way, including the sample size and R-squared.
Interpret the slope coefficient. If each mailing costs one guilder, is the charity expected to make a
net gain on each mailing? Does this mean the charity makes a net gain on every mailing? Explain.
What is the smallest predicted charitable contribution in the sample? Using this simple regression analysis, can you ever predict zero for gift?
C8 To complete this exercise you need a software package that allows you to generate data from the uniform and normal distributions.
(i) Start by generating 500 observations on xi—the explanatory variable—from the uniform distribution with range [0,10]. (Most statistical packages have a command for the Uniform(0,1)
distribution; just multiply those observations by 10.) What are the sample mean and sample
standard deviation of the xi?
58860_ch02_hr_019-065.indd 63
10/18/18 4:06 PM
64
PART 1 Regression Analysis with Cross-Sectional Data
(ii)
Randomly generate 500 errors, ui, from the Normal(0,36) distribution. (If you generate a
­Normal(0,1), as is commonly available, simply multiply the outcomes by six.) Is the sample
average of the ui exactly zero? Why or why not? What is the sample standard deviation of the ui?
(iii) Now generate the yi as
yi 5 1 1 2xi 1 ui ; b0 1 b1xi 1 ui;
that is, the population intercept is one and the population slope is two. Use the data to run the
regression of yi on xi. What are your estimates of the intercept and slope? Are they equal to the
population values in the above equation? Explain.
(iv) Obtain the OLS residuals, u^ i, and verify that equation (2.60) holds (subject to rounding error).
(v) Compute the same quantities in equation (2.60) but use the errors ui in place of the residuals.
Now what do you conclude?
(vi) Repeat parts (i), (ii), and (iii) with a new sample of data, starting with generating the xi. Now
what do you obtain for b^ 0 and b^ 1? Why are these different from what you obtained in part (iii)?
C9 Use the data in COUNTYMURDERS to answer these questions. Use only the data for 1996.
(i) How many counties had zero murders in 1996? How many counties had at least one execution?
What is the largest number of executions?
(ii) Estimate the equation
murders 5 b0 1 b1execs 1 u
by OLS and report the results in the usual way, including sample size and R-squared.
(iii) Interpret the slope coefficient reported in part (ii). Does the estimated equation suggest a deterrent effect of capital punishment?
(iv) What is the smallest number of murders that can be predicted by the equation? What is the
­residual for a county with zero executions and zero murders?
(v) Explain why a simple regression analysis is not well suited for determining whether capital punishment has a deterrent effect on murders.
C10 The data set in CATHOLIC includes test score information on over 7,000 students in the United States
who were in eighth grade in 1988. The variables math12 and read12 are scores on twelfth grade standardized math and reading tests, respectively.
(i) How many students are in the sample? Find the means and standard deviations of math12 and
read12.
(ii) Run the simple regression of math12 on read12 to obtain the OLS intercept and slope estimates.
Report the results in the form
math12 5 b^ 0 1 b^ 1read12
n 5 ?, R2 5 ?
where you fill in the values for b^ 0 and b^ 1 and also replace the question marks.
(iii) Does the intercept reported in part (ii) have a meaningful interpretation? Explain.
(iv) Are you surprised by the b^ 1 that you found? What about R2?
(v) Suppose that you present your findings to a superintendent of a school district, and the
superintendent says, “Your findings show that to improve math scores we just need to
improve reading scores, so we should hire more reading tutors.” How would you respond
to this comment? (Hint: If you instead run the regression of read12 on math12, what would
you expect to find?)
C11 Use the data in GPA1 to answer these questions. It is a sample of Michigan State University undergraduates from the mid-1990s, and includes current college GPA, colGPA, and a binary variable indicating
whether the student owned a personal computer (PC).
(i) How many students are in the sample? Find the average and highest college GPAs.
58860_ch02_hr_019-065.indd 64
10/18/18 4:06 PM
CHAPTER 2 The Simple Regression Model
65
(ii) How many students owned their own PC?
(iii) Estimate the simple regression equation
colGPA 5 b0 1 b1 PC 1 u
(iv)
(v)
and report your estimates for b0 and b1. Interpret these estimates, including a discussion ofthe
magnitudes.
What is the R-squared from the regression? What do you make of its magnitude?
Does your finding in part (iii) imply that owning a PC has a causal effect on colGPA? Explain.
Appendix 2A
Minimizing the Sum of Squared Residuals
We show that the OLS estimates b^ 0 and b^ 1 do minimize the sum of squared residuals, as asserted
in Section 2-2. Formally, the problem is to characterize the solutions b^ 0 and b^ 1 to the minimization
problem
min a 1 yi 2 b0 2 b1xi 2 2,
b0,b1
n
i51
where b0 and b1 are the dummy arguments for the optimization problem; for simplicity, call this
function Q 1 b0, b1 2 . By a fundamental result from multivariable calculus (see Math Refresher A), a
necessary condition for b^ 0 and b^ 1 to solve the minimization problem is that the partial derivatives of
^ 1: 'Q 1 b^ 0, b
^ 1 2 /'b0 5 0 and
Q 1 b0, b1 2 with respect to b0 and b1 must be zero when evaluated at b^ 0, b
^
^
1
2
'Q b0, b1 /'b1 5 0. Using the chain rule from calculus, these two equations become
22 a 1 yi 2 b^ 0 2 b^ 1xi 2 5 0
n
i51
22 a xi 1 yi 2 b^ 0 2 b^ 1xi 2 5 0.
n
i51
These two equations are just (2.14) and (2.15) multiplied by 22n and, therefore, are solved by the
same b^ 0 and b^ 1.
How do we know that we have actually minimized the sum of squared residuals? The first order
conditions are necessary but not sufficient conditions. One way to verify that we have minimized the
sum of squared residuals is to write, for any b0 and b1,
Q 1 b0, b1 2 5 a 3 yi 2 b^ 0 2 b^ 1xi 1 1 b^ 0 2 b0 2 1 1 b^ 1 2 b1 2 xi 4 2
n
i51
n
5 a 3 u^ i 1 1 b^ 0 2 b0 2 1 1 b^ 1 2 b1 2 xi 4 2
i51
n
n
n
i51
i51
i51
5 a u^ 2i 1 n 1 b^ 0 2 b0 2 2 1 1 b^ 1 2 b1 2 2 a x2i 1 2 1 b^ 0 2 b0 2 1 b^ 1 2 b1 2 a xi,
where we have used equations (2.30) and (2.31). The first term does not depend on b0 or b1, while
the sum of the last three terms can be written as
2
^
^
a 3 1 b0 2 b0 2 1 1 b1 2 b1 2 xi 4 ,
n
i51
as can be verified by straightforward algebra. Because this is a sum of squared terms, the
smallest it can be is zero. Therefore, it is smallest when b0 5 b^ 0 and b1 5 b^ 1.
58860_ch02_hr_019-065.indd 65
10/18/18 4:06 PM
chapter
3
Multiple Regression
Analysis: Estimation
I
n Chapter 2, we learned how to use simple regression analysis to explain a dependent variable, y,
as a function of a single independent variable, x. The primary drawback in using simple regression
analysis for empirical work is that it is very difficult to draw ceteris paribus conclusions about how
x affects y: the key assumption, SLR.4—that all other factors affecting y are uncorrelated with x—is
often unrealistic.
Multiple regression analysis is more amenable to ceteris paribus analysis because it allows us to
explicitly control for many other factors that simultaneously affect the dependent variable. This is important
both for testing economic theories and for evaluating policy effects when we must rely on nonexperimental
data. Because multiple regression models can accommodate many explanatory variables that may be correlated, we can hope to infer causality in cases where simple regression analysis would be misleading.
Naturally, if we add more factors to our model that are useful for explaining y, then more of the
variation in y can be explained. Thus, multiple regression analysis can be used to build better models
for predicting the dependent variable.
An additional advantage of multiple regression analysis is that it can incorporate fairly general
functional form relationships. In the simple regression model, only one function of a single explanatory variable can appear in the equation. As we will see, the multiple regression model allows for
much more flexibility.
Section 3-1 formally introduces the multiple regression model and further discusses the advantages of multiple regression over simple regression. In Section 3-2, we demonstrate how to estimate the parameters in the multiple regression model using the method of ordinary least squares.
66
58860_ch03_hr_066-116.indd 66
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
67
In Sections 3-3, 3-4, and 3-5, we describe various statistical properties of the OLS estimators, including unbiasedness and efficiency.
The multiple regression model is still the most widely used vehicle for empirical analysis in economics and other social sciences. Likewise, the method of ordinary least squares is popularly used for
estimating the parameters of the multiple regression model.
3-1 Motivation for Multiple Regression
3-1a The Model with Two Independent Variables
We begin with some simple examples to show how multiple regression analysis can be used to solve
problems that cannot be solved by simple regression.
The first example is a simple variation of the wage equation introduced in Chapter 2 for obtaining
the effect of education on hourly wage:
wage 5 b0 1 b1educ 1 b2exper 1 u,
[3.1]
where exper is years of labor market experience. Thus, wage is determined by the two explanatory
or independent variables, education and experience, and by other unobserved factors, which are contained in u. We are still primarily interested in the effect of educ on wage, holding fixed all other factors affecting wage; that is, we are interested in the parameter b1.
Compared with a simple regression analysis relating wage to educ, equation (3.1) effectively
takes exper out of the error term and puts it explicitly in the equation. Because exper appears in the equation,
its coefficient, b2, measures the ceteris paribus effect of exper on wage, which is also of some interest.
Not surprisingly, just as with simple regression, we will have to make assumptions about how u in
(3.1) is related to the independent variables, educ and exper. However, as we will see in Section 3-2,
there is one thing of which we can be confident: because (3.1) contains experience explicitly, we will
be able to measure the effect of education on wage, holding experience fixed. In a simple regression
analysis—which puts exper in the error term—we would have to assume that experience is uncorrelated with education, a tenuous assumption.
As a second example, consider the problem of explaining the effect of per-student spending
(expend) on the average standardized test score (avgscore) at the high school level. Suppose that the
average test score depends on funding, average family income (avginc), and other unobserved factors:
avgscore 5 b0 1 b1expend 1 b2avginc 1 u.
[3.2]
The coefficient of interest for policy purposes is b1, the ceteris paribus effect of expend on avgscore.
By including avginc explicitly in the model, we are able to control for its effect on avgscore. This is
likely to be important because average family income tends to be correlated with per-student spending: spending levels are often determined by both property and local income taxes. In simple regression analysis, avginc would be included in the error term, which would likely be correlated with
expend, causing the OLS estimator of b1 in the two-variable model to be biased.
In the two previous similar examples, we have shown how observable factors other than the variable of primary interest [educ in equation (3.1) and expend in equation (3.2)] can be included in a
regression model. Generally, we can write a model with two independent variables as
y 5 b0 1 b1x1 1 b2x2 1 u,
[3.3]
where
b0 is the intercept.
b1 measures the change in y with respect to x1, holding other factors fixed.
b2 measures the change in y with respect to x2, holding other factors fixed.
58860_ch03_hr_066-116.indd 67
10/18/18 4:22 PM
68
PART 1 Regression Analysis with Cross-Sectional Data
Multiple regression analysis is also useful for generalizing functional relationships between
variables. As an example, suppose family consumption (cons) is a quadratic function of family
income (inc):
cons 5 b0 1 b1inc 1 b2inc2 1 u,
[3.4]
where u contains other factors affecting consumption. In this model, consumption depends on only
one observed factor, income; so it might seem that it can be handled in a simple regression framework. But the model falls outside simple regression because it contains two functions of income, inc
and inc2 (and therefore three parameters, b0, b1, and b2). Nevertheless, the consumption function is
easily written as a regression model with two independent variables by letting x1 5 inc and x2 5 inc2.
Mechanically, there will be no difference in using the method of ordinary least squares (introduced in Section 3-2) to estimate equations as different as (3.1) and (3.4). Each equation can be
written as (3.3), which is all that matters for computation. There is, however, an important difference in how one interprets the parameters. In equation (3.1), b1 is the ceteris paribus effect of educ
on wage. The parameter b1 has no such interpretation in (3.4). In other words, it makes no sense to
measure the effect of inc on cons while holding inc2 fixed, because if inc changes, then so must inc2!
Instead, the change in consumption with respect to the change in income—the marginal propensity to
­consume—is approximated by
Dcons
< b1 1 2b2inc.
Dinc
See Math Refresher A for the calculus needed to derive this equation. In other words, the marginal
effect of income on consumption depends on b2 as well as on b1 and the level of income. This example shows that, in any particular application, the definitions of the independent variables are crucial.
But for the theoretical development of multiple regression, we can be vague about such details. We
will study examples like this more completely in Chapter 6.
In the model with two independent variables, the key assumption about how u is related to
x1 and x2 is
E 1 u 0 x1, x2 2 5 0.
[3.5]
The interpretation of condition (3.5) is similar to the interpretation of Assumption SLR.4 for simple
regression analysis. It means that, for any values of x1 and x2 in the population, the average of the
unobserved factors is equal to zero. As with simple regression, the important part of the assumption
is that the expected value of u is the same for all combinations of x1 and x2; that this common value
is zero is no assumption at all as long as the intercept b0 is included in the model (see Section 2-1).
How can we interpret the zero conditional mean assumption in the previous examples? In equation (3.1), the assumption is E 1 u 0 educ,exper 2 5 0. This implies that other factors affecting wage are
not related on average to educ and exper. Therefore,
if we think innate ability is part of u, then we will
G o i n g F u rt h e r 3 . 1
need average ability levels to be the same across all
A simple model to explain city murder rates
combinations of education and experience in the
(murdrate) in terms of the probability of
working population. This may or may not be true,
conviction (prbconv) and average sentence
but, as we will see in Section 3-3, this is the queslength (avgsen) is
tion we need to ask in order to determine whether the
murdrate 5 b0 1 b1prbconv
method of ordinary least squares produces unbiased
1 b2avgsen 1 u.
estimators.
The example measuring student performance
What are some factors contained in u?
[equation (3.2)] is similar to the wage equaDo you think the key assumption (3.5) is
tion. The zero conditional mean assumption is
likely to hold?
E 1 u 0 expend,avginc 2 5 0, which means that other
58860_ch03_hr_066-116.indd 68
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
69
factors affecting test scores—school or student characteristics—are, on average, unrelated to perstudent funding and average family income.
When applied to the quadratic consumption function in (3.4), the zero conditional mean assumption has a slightly different interpretation. Written literally, equation (3.5) becomes E 1 u 0 inc,inc2 2 5 0.
Because inc2 is known when inc is known, including inc2 in the expectation is redundant:
E 1 u 0 inc,inc2 2 5 0 is the same as E 1 u 0 inc 2 5 0. Nothing is wrong with putting inc2 along with inc in
the expectation when stating the assumption, but E 1 u 0 inc 2 5 0 is more concise.
3-1b The Model with k Independent Variables
Once we are in the context of multiple regression, there is no need to stop with two independent variables. Multiple regression analysis allows many observed factors to affect y. In the wage example, we
might also include amount of job training, years of tenure with the current employer, measures of ability, and even demographic variables like the number of siblings or mother’s education. In the school
funding example, additional variables might include measures of teacher quality and school size.
The general multiple linear regression (MLR) model (also called the multiple regression
model) can be written in the population as
y 5 b0 1 b1x1 1 b2x2 1 b3x3 1 p 1 bkxk 1 u,
[3.6]
where
b0 is the intercept.
b1 is the parameter associated with x1.
b2 is the parameter associated with x2, and so on.
Because there are k independent variables and an intercept, equation (3.6) contains k + 1 (unknown)
population parameters. For shorthand purposes, we will sometimes refer to the parameters other
than the intercept as slope parameters, even though this is not always literally what they are. [See
­equation (3.4), where neither b1 nor b2 is itself a slope, but together they determine the slope of the
relationship between consumption and income.]
The terminology for multiple regression is similar to that for simple regression and is given in
Table 3.1. Just as in simple regression, the variable u is the error term or disturbance. It contains
factors other than x1, x2, c, xk that affect y. No matter how many explanatory variables we include
in our model, there will always be factors we cannot include, and these are collectively contained in u.
When applying the general multiple regression model, we must know how to interpret the parameters. We will get plenty of practice now and in subsequent chapters, but it is useful at this point to be
reminded of some things we already know. Suppose that CEO salary (salary) is related to firm sales
(sales) and CEO tenure (ceoten) with the firm by
log 1 salary 2 5 b0 1 b1log 1 sales 2 1 b2ceoten 1 b3ceoten2 1 u.
[3.7]
This fits into the multiple regression model (with k 5 3) by defining y 5 log(salary), x1 5 log 1 sales 2 ,
x2 5 ceoten, and x3 5 ceoten2. As we know from Chapter 2, the parameter b1 is the ceteris paribus
Table 3.1 Terminology for Multiple Regression
58860_ch03_hr_066-116.indd 69
y
x1, x2, c, xk
Dependent variable
Independent variables
Explained variable
Explanatory variables
Response variable
Control variables
Predicted variable
Predictor variables
Regressand
Regressors
10/18/18 4:22 PM
70
PART 1 Regression Analysis with Cross-Sectional Data
elasticity of salary with respect to sales. If b3 5 0, then 100b2 is approximately the ceteris
paribus percentage increase in salary when ceoten increases by one year. When b3 2 0, the
effect of ceoten on salary is more complicated. We will postpone a detailed treatment of general models with quadratics until Chapter 6.
Equation (3.7) provides an important reminder about multiple regression analysis. The term “linear” in a multiple linear regression model means that equation (3.6) is linear in the parameters, bj.
Equation (3.7) is an example of a multiple regression model that, while linear in the bj, is a nonlinear
relationship between salary and the variables sales and ceoten. Many applications of multiple linear
regression involve nonlinear relationships among the underlying variables.
The key assumption for the general multiple regression model is easy to state in terms of a conditional expectation:
E 1 u 0 x1, x2, . . . , xk 2 5 0.
[3.8]
At a minimum, equation (3.8) requires that all factors in the unobserved error term be uncorrelated
with the explanatory variables. It also means that we have correctly accounted for the functional relationships between the explained and explanatory variables. Any problem that causes u to be correlated
with any of the independent variables causes (3.8) to fail. In Section 3-3, we will show that assumption (3.8) implies that OLS is unbiased and will derive the bias that arises when a key variable has
been omitted from the equation. In Chapters 15 and 16, we will study other reasons that might cause
(3.8) to fail and show what can be done in cases where it does fail.
3-2 Mechanics and Interpretation of Ordinary Least Squares
We now summarize some computational and algebraic features of the method of ordinary least squares
as it applies to a particular set of data. We also discuss how to interpret the estimated equation.
3-2a Obtaining the OLS Estimates
We first consider estimating the model with two independent variables. The estimated OLS equation
is written in a form similar to the simple regression case:
y^ 5 b^ 0 1 b^ 1x1 1 b^ 2x2,
[3.9]
where
b^ 0 = the estimate of b0.
b^ 1 = the estimate of b1.
b^ 2 = the estimate of b2.
But how do we obtain b^ 0, b^ 1, and b^ 2? The method of ordinary least squares chooses the estimates to minimize the sum of squared residuals. That is, given n observations on y, x1, and x2,
5 1 xi1, xi2, yi 2 : i 5 1, 2, . . . , n 6 , the estimates b^ 0, b^ 1, and b^ 2 are chosen simultaneously to make
2
^
^
^
a 1 yi 2 b0 2 b1xi1 2 b2xi2 2
n
i51
[3.10]
as small as possible.
To understand what OLS is doing, it is important to master the meaning of the indexing of the
independent variables in (3.10). The independent variables have two subscripts here, i followed by
either 1 or 2. The i subscript refers to the observation number. Thus, the sum in (3.10) is over all
i 5 1 to n observations. The second index is simply a method of distinguishing between different
independent variables. In the example relating wage to educ and exper, xi1 5 educi is education
for person i in the sample and xi2 5 experi is experience for person i. The sum of squared residun
als in equation (3.10) is g i51 1 wagei 2 b^ 0 2 b^ 1educi 2 b^ 2experi 2 2. In what follows, the i subscript
58860_ch03_hr_066-116.indd 70
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
71
is reserved for indexing the observation number. If we write xij, then this means the ith observation
on the jth independent variable. (Some authors prefer to switch the order of the observation number and the variable number, so that x1i is observation i on variable one. But this is just a matter of
notational taste.)
In the general case with k independent variables, we seek estimates, b^ 0, b^ 1, . . . , b^ k in the equation
y^ 5 b^ 0 1 b^ 1x1 1 b^ 2x2 1 p 1 b^ kxk.
[3.11]
The OLS estimates, k + 1 of them, are chosen to minimize the sum of squared residuals:
p 2 b^ kxik 2 2.
^
^
a 1 yi 2 b0 2 b1xi1 2
n
i51
[3.12]
This minimization problem can be solved using multivariable calculus (see Appendix 3A). This leads
to k + 1 linear equations in k + 1 unknowns b^ 0, b^ 1, . . . , b^ k:
p 2 b^ kxik 2 5 0
^
^
a 1 yi 2 b0 2 b1xi1 2
n
i51
p 2 b^ kxik 2 5 0
^
^
a xi1 1 yi 2 b0 2 b1xi1 2
n
i51
n
p 2 b^ kxik 2 5 0
^
^
a xi2 1 yi 2 b0 2 b1xi1 2
[3.13]
i51
(
p 2 b^ kxik 2 5 0.
^
^
a xik 1 yi 2 b0 2 b1xi1 2
n
i51
These are often called the OLS first order conditions. As with the simple regression model in
Section 2-2, the OLS first order conditions can be obtained by the method of moments: under assumption (3.8), E 1 u 2 5 0 and E 1 xju 2 5 0, where j 5 1, 2, . . . , k. The equations in (3.13) are the sample
counterparts of these population moments, although we have omitted the division by the sample
size n.
For even moderately sized n and k, solving the equations in (3.13) by hand calculations is tedious.
Nevertheless, modern computers running standard statistics and econometrics software can solve
these equations with large n and k very quickly.
There is only one slight caveat: we must assume that the equations in (3.13) can be solved
uniquely for the b^ j. For now, we just assume this, as it is usually the case in well-specified models.
In Section 3-3, we state the assumption needed for unique OLS estimates to exist (see Assumption
MLR.3).
As in simple regression analysis, equation (3.11) is called the OLS regression line or the sample
regression function (SRF). We will call b^ 0 the OLS intercept estimate and b^ 1, . . . , b^ k the OLS
slope estimates (corresponding to the independent variables x1, x2, . . . , xk).
To indicate that an OLS regression has been run, we will either write out equation (3.11) with y
and x1, . . . , xk replaced by their variable names (such as wage, educ, and exper), or we will say that
“we ran an OLS regression of y on x1, x2, . . . , xk” or that “we regressed y on x1, x2, . . . , xk.” These
are shorthand for saying that the method of ordinary least squares was used to obtain the OLS
equation (3.11). Unless explicitly stated otherwise, we always estimate an intercept along with the
slopes.
3-2b Interpreting the OLS Regression Equation
More important than the details underlying the computation of the b^ j is the interpretation of the
estimated equation. We begin with the case of two independent variables:
y^ 5 b^ 0 1 b^ 1x1 1 b^ 2x2.
58860_ch03_hr_066-116.indd 71
[3.14]
10/18/18 4:22 PM
72
PART 1 Regression Analysis with Cross-Sectional Data
The intercept b^ 0 in equation (3.14) is the predicted value of y when x1 5 0 and x2 5 0. Sometimes,
setting x1 and x2 both equal to zero is an interesting scenario; in other cases, it will not make sense.
Nevertheless, the intercept is always needed to obtain a prediction of y from the OLS regression line,
as (3.14) makes clear.
The estimates b^ 1 and b^ 2 have partial effect, or ceteris paribus, interpretations. From equation
(3.14), we have
Dy^ 5 b^ 1Dx1 1 b^ 2Dx2,
so we can obtain the predicted change in y given the changes in x1 and x2. (Note how the intercept has
nothing to do with the changes in y.) In particular, when x2 is held fixed, so that Dx2 5 0, then
Dy^ 5 b^ 1Dx1,
holding x2 fixed. The key point is that, by including x2 in our model, we obtain a coefficient on x1 with
a ceteris paribus interpretation. This is why multiple regression analysis is so useful. Similarly,
Dy^ 5 b^ 2Dx2,
holding x1 fixed.
Example 3.1
Determinants of College GPA
The variables in GPA1 include the college grade point average (colGPA), high school GPA (hsGPA),
and achievement test score (ACT) for a sample of 141 students from a large university; both college
and high school GPAs are on a four-point scale. We obtain the following OLS regression line to predict college GPA from high school GPA and achievement test score:
colGPA 5 1.29 1 .453 hsGPA 1 .0094 ACT
n 5 141.
[3.15]
How do we interpret this equation? First, the intercept 1.29 is the predicted college GPA if hsGPA and
ACT are both set as zero. Because no one who attends college has either a zero high school GPA or a
zero on the achievement test, the intercept in this equation is not, by itself, meaningful.
More interesting estimates are the slope coefficients on hsGPA and ACT. As expected, there is a positive
partial relationship between colGPA and hsGPA: holding ACT fixed, another point on hsGPA is associated
with .453 of a point on the college GPA, or almost half a point. In other words, if we choose two students,
A and B, and these students have the same ACT score, but the high school GPA of Student A is one point
higher than the high school GPA of Student B, then we predict Student A to have a college GPA .453 higher
than that of Student B. (This says nothing about any two actual people, but it is our best prediction.)
The sign on ACT implies that, while holding hsGPA fixed, a change in the ACT score of 10 points—
a very large change, as the maximum ACT score is 36 and the average score in the sample is about 24
with a standard deviation less than three—affects colGPA by less than one-tenth of a point. This is a
small effect, and it suggests that, once high school GPA is accounted for, the ACT score is not a strong
predictor of college GPA. (Naturally, there are many other factors that contribute to GPA, but here we
focus on statistics available for high school students.) Later, after we discuss statistical inference, we
will show that not only is the coefficient on ACT practically small, it is also statistically insignificant.
If we focus on a simple regression analysis relating colGPA to ACT only, we obtain
colGPA 5 2.40 1 .0271 ACT
n 5 141;
thus, the coefficient on ACT is almost three times as large as the estimate in (3.15). But this equation
does not allow us to compare two people with the same high school GPA; it corresponds to a different
experiment. We say more about the differences between multiple and simple regression later.
58860_ch03_hr_066-116.indd 72
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
73
The case with more than two independent variables is similar. The OLS regression line is
y^ 5 b^ 0 1 b^ 1x1 1 b^ 2x2 1 p 1 b^ kxk.
[3.16]
Dy^ 5 b^ 1Dx1 1 b^ 2Dx2 1 p 1 b^ kDxk.
[3.17]
Written in terms of changes,
The coefficient on x1 measures the change in y^ due to a one-unit increase in x1, holding all other independent variables fixed. That is,
Dy^ 5 b^ 1Dx1,
[3.18]
holding x2, x3, . . . , xk fixed. Thus, we have controlled for the variables x2, x3, . . . , xk when estimating
the effect of x1 on y. The other coefficients have a similar interpretation.
The following is an example with three independent variables.
Example 3.2
Hourly Wage Equation
Using the 526 observations on workers in WAGE1, we include educ (years of education), exper (years
of labor market experience), and tenure (years with the current employer) in an equation explaining
log(wage). The estimated equation is
log 1 wage 2 5 .284 1 .092 educ 1 .0041 exper 1 .022 tenure
n 5 526.
[3.19]
As in the simple regression case, the coefficients have a percentage interpretation. The only difference
here is that they also have a ceteris paribus interpretation. The coefficient .092 means that, holding
exper and tenure fixed, another year of education is predicted to increase log(wage) by .092, which
translates into an approximate 9.2% [100(.092)] increase in wage. Alternatively, if we take two people
with the same levels of experience and job tenure, the coefficient on educ is the proportionate difference in predicted wage when their education levels differ by one year. This measure of the return
to education at least keeps two important productivity factors fixed; whether it is a good estimate of
the ceteris paribus return to another year of education requires us to study the statistical properties of
OLS (see Section 3-3).
3-2c On the Meaning of “Holding Other Factors Fixed”
in Multiple Regression
The partial effect interpretation of slope coefficients in multiple regression analysis can cause some
confusion, so we provide a further discussion now.
In Example 3.1, we observed that the coefficient on ACT measures the predicted difference in
colGPA, holding hsGPA fixed. The power of multiple regression analysis is that it provides this ceteris
paribus interpretation even though the data have not been collected in a ceteris paribus fashion. In giving the coefficient on ACT a partial effect interpretation, it may seem that we actually went out and
sampled people with the same high school GPA but possibly with different ACT scores. This is not the
case. The data are a random sample from a large university: there were no restrictions placed on the
sample values of hsGPA or ACT in obtaining the data. Rarely do we have the luxury of holding certain
variables fixed in obtaining our sample. If we could collect a sample of individuals with the same high
school GPA, then we could perform a simple regression analysis relating colGPA to ACT. Multiple
regression effectively allows us to mimic this situation without restricting the values of any independent variables.
58860_ch03_hr_066-116.indd 73
10/18/18 4:22 PM
74
PART 1 Regression Analysis with Cross-Sectional Data
The power of multiple regression analysis is that it allows us to do in nonexperimental environments what natural scientists are able to do in a controlled laboratory setting: keep other factors fixed.
3-2d Changing More Than One Independent Variable
Simultaneously
Sometimes, we want to change more than one independent variable at the same time to find the resulting effect on the dependent variable. This is easily done using equation (3.17). For example, in equation (3.19), we can obtain the estimated effect on wage when an individual stays at the same firm for
another year: exper (general workforce experience) and tenure both increase by one year. The total
effect (holding educ fixed) is
Dlog 1 wage 2 5 .0041 Dexper 1 .022 Dtenure 5 .0041 1 .022 5 .0261,
or about 2.6%. Because exper and tenure each increase by one year, we just add the coefficients on
exper and tenure and multiply by 100 to turn the effect into a percentage.
3-2e OLS Fitted Values and Residuals
After obtaining the OLS regression line (3.11), we can obtain a fitted or predicted value for each
observation. For observation i, the fitted value is simply
y^ i 5 b^ 0 1 b^ 1xi1 1 b^ 2xi2 1 p 1 b^ kxik,
[3.20]
which is just the predicted value obtained by plugging the values of the independent variables for
observation i into equation (3.11). We should not forget about the intercept in obtaining the fitted
values; otherwise, the answer can be very misleading. As an example, if in (3.15), hsGPAi 5 3.5 and
ACTi 5 24, colGPAi 5 1.29 1 .453 1 3.5 2 1 .0094 1 24 2 5 3.101 (rounded to three places after the
decimal).
Normally, the actual value yi for any observation i will not equal the predicted value, y^ i :
OLS minimizes the average squared prediction error, which says nothing about the prediction
error for any particular observation. The residual for observation i is defined just as in the simple
regression case,
u^ i 5 yi 2 y^ i.
[3.21]
There is a residual for each observation. If u^ i . 0, then y^ i is below yi, which means that, for this
observation, yi is underpredicted. If u^ i , 0, then yi , y^ i, and yi is overpredicted.
The OLS fitted values and residuals have some important properties that are immediate extensions from the single variable case:
1. The sample average of the residuals is zero and so y 5 y^ .
G o i n g F u rt h e r 3 . 2
In Example 3.1, the OLS fitted line explaining college GPA in terms of high school GPA
and ACT score is
colGPA 5 1.29 1 .453 hsGPA
1 .0094 ACT.
If the average high school GPA is about
3.4 and the average ACT score is about
24.2, what is the average college GPA in the
sample?
58860_ch03_hr_066-116.indd 74
2. The sample covariance between each independent variable and
the OLS residuals is zero. Consequently, the sample covariance
between the OLS fitted values and the OLS residuals is zero.
3. The point 1 x1, x2, . . . , xk, y 2 is always on the OLS regression
line: y 5 b^ 0 1 b^ 1x1 1 b^ 2x2 1 p 1 b^ kxk.
The first two properties are immediate consequences of the set
of equations used to obtain the OLS estimates. The first equation in
(3.13) says that the sum of the residuals is zero. The remaining equan
tions are of the form g i51 xiju^ i 5 0, which implies that each independent variable has zero sample covariance with u^ i. Property (3)
follows immediately from property (1).
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
75
3-2f A “Partialling Out” Interpretation of Multiple Regression
When applying OLS, we do not need to know explicit formulas for the b^ j that solve the system of
equations in (3.13). Nevertheless, for certain derivations, we do need explicit formulas for the b^ j.
These formulas also shed further light on the workings of OLS.
Consider again the case with k 5 2 independent variables, y^ 5 b^ 0 1 b^ 1x1 1 b^ 2x2. For concreteness, we focus on b^ 1. One way to express b^ 1 is
b^ 1 5 a a r^ i1yi b ^ a a r^ 2i1 b,
n
n
i51
i51
[3.22]
where the r^ i1 are the OLS residuals from a simple regression of x1 on x2, using the sample at hand. We
regress our first independent variable, x1, on our second independent variable, x2, and then obtain the
residuals (y plays no role here). Equation (3.22) shows that we can then do a simple regression of y
on r^ 1 to obtain b^ 1. (Note that the residuals r^ i1 have a zero sample average, and so b^ 1 is the usual slope
estimate from simple regression.)
The representation in equation (3.22) gives another demonstration of b^ 1’s partial effect interpretation. The residuals r^ i1 are the part of xi1 that is uncorrelated with xi2. Another way of saying this
is that r^ i1 is xi1 after the effects of xi2 have been partialled out, or netted out. Thus, b^ 1 measures the
sample relationship between y and x1 after x2 has been partialled out.
In simple regression analysis, there is no partialling out of other variables because no other variables are included in the regression. Computer Exercise C5 steps you through the partialling out process using the wage data from Example 3.2. For practical purposes, the important thing is that b^ 1 in
the equation y^ 5 b^ 0 1 b^ 1x1 1 b^ 2x2 measures the change in y given a one-unit increase in x1, holding
x2 fixed.
In the general model with k explanatory variables, b^ 1 can still be written as in equation (3.22), but
the residuals r^ i1 come from the regression of x1 on x2, . . . , xk. Thus, b^ 1 measures the effect of x1 on y
after x2, . . . , xk have been partialled or netted out. In econometrics, the general partialling out result is
usually called the Frisch-Waugh theorem. It has many uses in theoretical and applied econometrics.
We will see applications to time series regressions in Chapter 10.
3-2g Comparison of Simple and Multiple Regression Estimates
Two special cases exist in which the simple regression of y on x1 will produce the same OLS estimate
on x1 as the regression of y on x1 and x2. To be more precise, write the simple regression of y on x1 as
| 1b
| x , and write the multiple regression as y^ 5 b^ 1 b^ x 1 b^ x . We know that the simple
|
y5b
0
1 1
0
1 1
2 2
| does not usually equal the multiple regression coefficient b^ . It turns out
regression coefficient b
1
1
| and b^ , which allows for interesting comparisons between
there is a simple relationship between b
1
1
simple and multiple regression:
| 5 b^ 1 b^ |
b
d,
[3.23]
1
1
2 1
where |
d 1 is the slope coefficient from the simple regression of xi2 on xi1, i 5 1, . . . , n. This equation
| differs from the partial effect of x on y^ . The confounding term is the partial effect of
shows how b
1
1
x2 on y^ times the slope in the simple regression of x2 on x1. (See Section 3A-4 in the chapter appendix
for a more general verification.)
| and b^ also shows there are two distinct cases where they are equal:
The relationship between b
1
1
1. The partial effect of x2 on y^ is zero in the sample. That is, b^ 2 5 0.
&
2. x1 and x2 are uncorrelated in the sample. That is, d1 5 0.
Even though simple and multiple regression estimates are almost never identical, we can use
the above formula to characterize why they might be either very different or quite similar. For example, if b^ 2 is small, we might expect the multiple and simple regression estimates of b1 to be similar.
58860_ch03_hr_066-116.indd 75
10/18/18 4:22 PM
76
PART 1 Regression Analysis with Cross-Sectional Data
In Example 3.1, the sample correlation between hsGPA and ACT is about .346, which is a nontrivial
correlation. But the coefficient on ACT is pretty small. Therefore, it is not surprising to find that the
simple regression of colGPA on hsGPA produces a slope estimate of .482, which is not much different
from the estimate .453 in (3.15).
Example 3.3
Participation in 401(k) Pension Plans
We use the data in 401K to estimate the effect of a plan’s match rate (mrate) on the participation rate
(prate) in its 401(k) pension plan. The match rate is the amount the firm contributes to a worker’s
fund for each dollar the worker contributes (up to some limit); thus, mrate 5 .75 means that the firm
contributes 75¢ for each dollar contributed by the worker. The participation rate is the percentage of
eligible workers having a 401(k) account. The variable age is the age of the 401(k) plan. There are
1,534 plans in the data set, the average prate is 87.36, the average mrate is .732, and the average age
is 13.2.
Regressing prate on mrate, age gives
prate 5 80.12 1 5.52 mrate 1 .243 age
n 5 1,534.
Thus, both mrate and age have the expected effects. What happens if we do not control for age?
The estimated effect of age is not trivial, and so we might expect a large change in the estimated
effect of mrate if age is dropped from the regression. However, the simple regression of prate on mrate
yields prate 5 83.08 1 5.86 mrate. The simple regression estimate of the effect of mrate on prate is
clearly different from the multiple regression estimate, but the difference is not very big. (The simple regression estimate is only about 6.2% larger than the multiple regression estimate.) This can be
explained by the fact that the sample correlation between mrate and age is only .12.
In the case with k independent variables, the simple regression of y on x1 and the multiple regression of y on x1, x2, c, xk produce an identical estimate of x1 only if (1) the OLS coefficients on x2
through xk are all zero or (2) x1 is uncorrelated with each of x2, c, xk. Neither of these is very likely
in practice. But if the coefficients on x2 through xk are small, or the sample correlations between x1
and the other independent variables are insubstantial, then the simple and multiple regression estimates of the effect of x1 on y can be similar.
3-2h Goodness-of-Fit
As with simple regression, we can define the total sum of squares (SST), the explained sum of
squares (SSE), and the residual sum of squares or sum of squared residuals (SSR) as
SST ; a 1 yi 2 y 2 2
n
i51
SSE ; a 1 y^ i 2 y 2 2
n
i51
SSR ; a u^ i2.
n
i51
[3.24]
[3.25]
[3.26]
Using the same argument as in the simple regression case, we can show that
SST 5 SSE 1 SSR.
[3.27]
In other words, the total variation in 5 yi 6 is the sum of the total variations in 5 y^ i 6 and in 5 u^ i 6 .
58860_ch03_hr_066-116.indd 76
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
77
Assuming that the total variation in y is nonzero, as is the case unless yi is constant in the sample,
we can divide (3.27) by SST to get
SSR/SST 1 SSE/SST 5 1.
Just as in the simple regression case, the R-squared is defined to be
R2 ; SSE/SST 5 1 2 SSR/SST,
[3.28]
and it is interpreted as the proportion of the sample variation in yi that is explained by the OLS regression line. By definition, R2 is a number between zero and one.
R2 can also be shown to equal the squared correlation coefficient between the actual yi and the
fitted values y^ i. That is,
a a 1 yi 2 y 2 1 y^ i 2 y^ 2 b
n
R2 5
2
i51
a a 1 yi 2 y 2 2 b a a 1 y^ i 2 y^ 2 2 b
n
i51
n
.
[3.29]
i51
[We have put the average of the y^ i in (3.29) to be true to the formula for a correlation coefficient;
we know that this average equals y because the sample average of the residuals is zero and
yi 5 y^ i 1 u^ i.]
Example 3.4
Determinants of College GPA
From the grade point average regression that we did earlier, the equation with R2 is
colGPA 5 1.29 1 .453 hsGPA 1 .0094 ACT
n 5 141, R2 5 .176.
This means that hsGPA and ACT together explain about 17.6% of the variation in college GPA
for this sample of students. This may not seem like a high percentage, but we must remember that
there are many other factors—including family background, personality, quality of high school
­education, affinity for college—that contribute to a student’s college performance. If hsGPA and ACT
explained almost all of the variation in colGPA, then performance in college would be preordained by
high school performance!
An important fact about R2 is that it never decreases, and it usually increases, when another independent variable is added to a regression and the same set of observations is used for both regressions.
This algebraic fact follows because, by definition, the sum of squared residuals never increases when
additional regressors are added to the model. For example, the last digit of one’s social security number has nothing to do with one’s hourly wage, but adding this digit to a wage equation will increase
the R2 (by a little, at least).
An important caveat to the previous assertion about R-squared is that it assumes we do not have
missing data on the explanatory variables. If two regressions use different sets of observations, then,
in general, we cannot tell how the R-squareds will compare, even if one regression uses a subset of
regressors. For example, suppose we have a full set of data on the variables y, x1, and x2, but for some
units in our sample data are missing on x3. Then we cannot say that the R-squared from regressing y
on x1, x2 will be less than that from regressing y on x1, x2, and x3: it could go either way. Missing data
can be an important practical issue, and we will return to it in Chapter 9.
The fact that R2 never decreases when any variable is added to a regression makes it a poor tool
for deciding whether one variable or several variables should be added to a model. The factor that
should determine whether an explanatory variable belongs in a model is whether the explanatory
58860_ch03_hr_066-116.indd 77
10/18/18 4:22 PM
78
PART 1 Regression Analysis with Cross-Sectional Data
variable has a nonzero partial effect on y in the population. We will show how to test this hypothesis in Chapter 4 when we cover statistical inference. We will also see that, when used properly, R2
allows us to test a group of variables to see if it is important for explaining y. For now, we use it as a
goodness-of-fit measure for a given model.
Example 3.5
Explaining Arrest Records
CRIME1 contains data on arrests during the year 1986 and other information on 2,725 men born in
either 1960 or 1961 in California. Each man in the sample was arrested at least once prior to 1986.
The variable narr86 is the number of times the man was arrested during 1986: it is zero for most
men in the sample (72.29%), and it varies from 0 to 12. (The percentage of men arrested once during
1986 was 20.51.) The variable pcnv is the proportion (not percentage) of arrests prior to 1986 that led
to conviction, avgsen is average sentence length served for prior convictions (zero for most people),
ptime86 is months spent in prison in 1986, and qemp86 is the number of quarters during which the
man was employed in 1986 (from zero to four).
A linear model explaining arrests is
narr86 5 b0 1 b1pcnv 1 b2avgsen 1 b3ptime86 1 b4qemp86 1 u,
where pcnv is a proxy for the likelihood for being convicted of a crime and avgsen is a measure of
expected severity of punishment, if convicted. The variable ptime86 captures the incarcerative effects
of crime: if an individual is in prison, he cannot be arrested for a crime outside of prison. Labor market opportunities are crudely captured by qemp86.
First, we estimate the model without the variable avgsen. We obtain
narr86 5 .712 2 .150 pcnv 2 .034 ptime86 2 .104 qemp86
n 5 2,725, R2 5 .0413.
This equation says that, as a group, the three variables pcnv, ptime86, and qemp86 explain about 4.1%
of the variation in narr86.
Each of the OLS slope coefficients has the anticipated sign. An increase in the proportion of
convictions lowers the predicted number of arrests. If we increase pcnv by .50 (a large increase in the
probability of conviction), then, holding the other factors fixed, Dnarr86 5 2.150 1 .50 2 5 2.075.
This may seem unusual because an arrest cannot change by a fraction. But we can use this value to
obtain the predicted change in expected arrests for a large group of men. For example, among 100 men,
the predicted fall in arrests when pcnv increases by .50 is 27.5.
Similarly, a longer prison term leads to a lower predicted number of arrests. In fact, if ptime86
increases from 0 to 12, predicted arrests for a particular man fall by .034 1 12 2 5 .408. Another quarter
in which legal employment is reported lowers predicted arrests by .104, which would be 10.4 arrests
among 100 men.
If avgsen is added to the model, we know that R2 will increase. The estimated equation is
narr86 5 .707 2 .151 pcnv 1 .0074 avgsen 2 .037 ptime86 2 .103 qemp86
n 5 2,725, R2 5 .0422.
Thus, adding the average sentence variable increases R2 from .0413 to .0422, a practically small
effect. The sign of the coefficient on avgsen is also unexpected: it says that a longer average sentence
length increases criminal activity.
Example 3.5 deserves a final word of caution. The fact that the four explanatory variables included
in the second regression explain only about 4.2% of the variation in narr86 does not necessarily
58860_ch03_hr_066-116.indd 78
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
79
mean that the equation is useless. Even though these variables collectively do not explain much of
the variation in arrests, it is still possible that the OLS estimates are reliable estimates of the ceteris
paribus effects of each independent variable on narr86. As we will see, whether this is the case does
not directly depend on the size of R2. Generally, a low R2 indicates that it is hard to predict individual
outcomes on y with much accuracy, something we study in more detail in Chapter 6. In the arrest
example, the small R2 reflects what we already suspect in the social sciences: it is generally very difficult to predict individual behavior.
3-2i Regression through the Origin
Sometimes, an economic theory or common sense suggests that b0 should be zero, and so we should
briefly mention OLS estimation when the intercept is zero. Specifically, we now seek an equation of
the form
|x 1b
|x 1p1b
|x,
|
y5b
[3.30]
1 1
2 2
k k
where the symbol “|” over the estimates is used to distinguish them from the OLS estimates obtained
along with the intercept [as in (3.11)]. In (3.30), when x1 5 0, x2 5 0, . . . , xk 5 0, the predicted
|,...,b
| are said to be the OLS estimates from the regression of y on
value is zero. In this case, b
1
k
x1, x2, . . . , xk through the origin.
The OLS estimates in (3.30), as always, minimize the sum of squared residuals, but with the
intercept set at zero. You should be warned that the properties of OLS that we derived earlier no
longer hold for regression through the origin. In particular, the OLS residuals no longer have a zero
sample average. Further, if R2 is defined as 1 2 SSR/SST, where SST is given in (3.24) and SSR is
n
|x 2p2b
| x 2 2, then R2 can actually be negative. This means that the sample
now a i51 1 yi 2 b
1 i1
k ik
average, y, “explains” more of the variation in the yi than the explanatory variables. Either we should
include an intercept in the regression or conclude that the explanatory variables poorly explain y. To
always have a nonnegative R-squared, some economists prefer to calculate R2 as the squared correlation coefficient between the actual and fitted values of y, as in (3.29). (In this case, the average fitted
value must be computed directly because it no longer equals y.) However, there is no set rule on computing R-squared for regression through the origin.
One serious drawback with regression through the origin is that, if the intercept b0 in the population model is different from zero, then the OLS estimators of the slope parameters will be biased. The
bias can be severe in some cases. The cost of estimating an intercept when b0 is truly zero is that the
variances of the OLS slope estimators are larger.
3-3 The Expected Value of the OLS Estimators
We now turn to the statistical properties of OLS for estimating the parameters in an underlying
population model. In this section, we derive the expected value of the OLS estimators. In particular, we state and discuss four assumptions, which are direct extensions of the simple regression
model assumptions, under which the OLS estimators are unbiased for the population parameters.
We also explicitly obtain the bias in OLS when an important variable has been omitted from the
regression.
You should remember that statistical properties have nothing to do with a particular sample, but
rather with the property of estimators when random sampling is done repeatedly. Thus, Sections 3-3, 3-4,
and 3-5 are somewhat abstract. Although we give examples of deriving bias for particular models, it
is not meaningful to talk about the statistical properties of a set of estimates obtained from a single
sample.
The first assumption we make simply defines the multiple linear regression (MLR) model.
58860_ch03_hr_066-116.indd 79
10/18/18 4:22 PM
80
PART 1 Regression Analysis with Cross-Sectional Data
Assumption MLR.1
Linear in Parameters
The model in the population can be written as
y 5 b0 1 b1x1 1 b2x2 1 p 1 bkxk 1 u,[3.31]
where b0, b1, . . . , bk are the unknown parameters (constants) of interest and u is an unobserved
random error or disturbance term.
Equation (3.31) formally states the population model, sometimes called the true model, to allow
for the possibility that we might estimate a model that differs from (3.31). The key feature is that the
model is linear in the parameters b0, b1, . . . , bk. As we know, (3.31) is quite flexible because y and
the independent variables can be arbitrary functions of the underlying variables of interest, such as
natural logarithms and squares [see, for example, equation (3.7)].
Assumption MLR.2
Random Sampling
We have a random sample of n observations, 5 1 xi1, xi2, . . . , xik, yi 2 : i 5 1, 2, . . . , n 6 , following the
population model in Assumption MLR.1.
Sometimes, we need to write the equation for a particular observation i: for a randomly drawn
observation from the population, we have
yi 5 b0 1 b1xi1 1 b2xi2 1 p 1 bkxik 1 ui.
[3.32]
Remember that i refers to the observation, and the second subscript on x is the variable number. For
example, we can write a CEO salary equation for a particular CEO i as
log 1 salaryi 2 5 b0 1 b1log 1 salesi 2 1 b2ceoteni 1 b3ceoten2i 1 ui.
[3.33]
The term ui contains the unobserved factors for CEO i that affect his or her salary. For applications,
it is usually easiest to write the model in population form, as in (3.31). It contains less clutter and
emphasizes the fact that we are interested in estimating a population relationship.
In light of model (3.31), the OLS estimators b^ 0, b^ 1, b^ 2, . . . , b^ k from the regression of y on
x1, . . . , xk are now considered to be estimators of b0, b1, . . . , bk. In Section 3-2, we saw that OLS
chooses the intercept and slope estimates for a particular sample so that the residuals average to zero
and the sample correlation between each independent variable and the residuals is zero. Still, we did
not include conditions under which the OLS estimates are well defined for a given sample. The next
assumption fills that gap.
Assumption MLR.3
No Perfect Collinearity
In the sample (and therefore in the population), none of the independent variables is constant, and there
are no exact linear relationships among the independent variables.
Assumption MLR.3 is more complicated than its counterpart for simple regression because we must
now look at relationships between all independent variables. If an independent variable in (3.31) is
an exact linear combination of the other independent variables, then we say the model suffers from
perfect collinearity, and it cannot be estimated by OLS.
It is important to note that Assumption MLR.3 does allow the independent variables to be correlated; they just cannot be perfectly correlated. If we did not allow for any correlation among the
58860_ch03_hr_066-116.indd 80
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
81
independent variables, then multiple regression would be of very limited use for econometric analysis.
For example, in the model relating test scores to educational expenditures and average family income,
avgscore 5 b0 1 b1expend 1 b2avginc 1 u,
we fully expect expend and avginc to be correlated: school districts with high average family incomes
tend to spend more per student on education. In fact, the primary motivation for including avginc
in the equation is that we suspect it is correlated with expend, and so we would like to hold it fixed
in the analysis. Assumption MLR.3 only rules out perfect correlation between expend and avginc in
our sample. We would be very unlucky to obtain a sample where per-student expenditures are perfectly correlated with average family income. But some correlation, perhaps a substantial amount, is
expected and certainly allowed.
The simplest way that two independent variables can be perfectly correlated is when one variable is a constant multiple of another. This can happen when a researcher inadvertently puts the same
variable measured in different units into a regression equation. For example, in estimating a relationship between consumption and income, it makes no sense to include as independent variables income
measured in dollars as well as income measured in thousands of dollars. One of these is redundant.
What sense would it make to hold income measured in dollars fixed while changing income measured
in thousands of dollars?
We already know that different nonlinear functions of the same variable can appear among
the regressors. For example, the model cons 5 b0 1 b1inc 1 b2inc2 1 u does not violate
Assumption MLR.3: even though x2 5 inc2 is an exact function of x1 5 inc, inc2 is not an exact ­linear
function of inc. Including inc2 in the model is a useful way to generalize functional form, unlike
including income measured in dollars and in thousands of dollars.
Common sense tells us not to include the same explanatory variable measured in different units
in the same regression equation. There are also more subtle ways that one independent variable can
be a multiple of another. Suppose we would like to estimate an extension of a constant elasticity consumption function. It might seem natural to specify a model such as
log 1 cons 2 5 b0 1 b1log 1 inc 2 1 b2log 1 inc2 2 1 u,
[3.34]
voteA 5 b0 1 b1expendA 1 b2expendB 1 b3totexpend 1 u,
[3.35]
where x1 5 log 1 inc 2 and x2 5 log 1 inc2 2 . Using the basic properties of the natural log (see Math
Refresher A), log 1 inc2 2 5 2 # log 1 inc 2 . That is, x2 5 2x1, and naturally this holds for all observations
in the sample. This violates Assumption MLR.3. What we should do instead is include 3 log 1 inc 2 4 2,
not log 1 inc2 2 , along with log(inc). This is a sensible extension of the constant elasticity model, and
we will see how to interpret such models in Chapter 6.
Another way that independent variables can be perfectly collinear is when one independent variable can be expressed as an exact linear function of two or more of the other independent variables.
For example, suppose we want to estimate the effect of campaign spending on campaign outcomes.
For simplicity, assume that each election has two candidates. Let voteA be the percentage of the
vote for Candidate A, let expendA be campaign expenditures by Candidate A, let expendB be campaign expenditures by Candidate B, and let totexpend be total campaign expenditures; the latter three
variables are all measured in dollars. It may seem natural to specify the model as
in order to isolate the effects of spending by each candidate and the total amount of spending. But this
model violates Assumption MLR.3 because x3 5 x1 1 x2 by definition. Trying to interpret this equation in a ceteris paribus fashion reveals the problem. The parameter of b1 in equation (3.35) is supposed to measure the effect of increasing expenditures by Candidate A by one dollar on Candidate A’s
vote, holding Candidate B’s spending and total spending fixed. This is nonsense, because if expendB
and totexpend are held fixed, then we cannot increase expendA.
The solution to the perfect collinearity in (3.35) is simple: drop any one of the three variables
from the model. We would probably drop totexpend, and then the coefficient on expendA would
58860_ch03_hr_066-116.indd 81
10/18/18 4:22 PM
82
PART 1 Regression Analysis with Cross-Sectional Data
measure the effect of increasing expenditures by A on the percentage of the vote received by A, holding the spending by B fixed.
The prior examples show that Assumption MLR.3 can fail if we are not careful in specifying
our model. Assumption MLR.3 also fails if the sample size, n, is too small in relation to the number
of parameters being estimated. In the general regression model in equation (3.31), there are k 1 1
parameters, and MLR.3 fails if n , k 1 1. Intuitively, this makes sense: to estimate k 1 1 parameters,
we need at least k 1 1 observations. Not surprisingly, it is better to have as many observations as possible, something we will see with our variance calculations in Section 3-4.
If the model is carefully specified and n $ k 1 1,
Assumption
MLR.3 can fail in rare cases due to
G o i n g F u rt h e r 3 . 3
bad luck in collecting the sample. For example, in
In the previous example, if we use as explanaa wage equation with education and experience as
tory variables expendA, expendB, and shareA,
variables, it is possible that we could obtain a ranwhere shareA 5 100 #(expendA/­totexpend)
dom sample where each individual has exactly twice
is the percentage share of total campaign
as much education as years of experience. This sceexpenditures made by Candidate A, does this
nario would cause Assumption MLR.3 to fail, but it
violate Assumption MLR.3?
can be considered very unlikely unless we have an
extremely small sample size.
The final, and most important, assumption needed for unbiasedness is a direct extension of
Assumption SLR.4.
Assumption MLR.4
Zero Conditional Mean
The error u has an expected value of zero given any values of the independent variables. In other words,
E 1 u 0 x1, x2, . . . , xk 2 5 0.[3.36]
One way that Assumption MLR.4 can fail is if the functional relationship between the explained and
explanatory variables is misspecified in equation (3.31): for example, if we forget to include the quadratic term inc2 in the consumption function cons 5 b0 1 b1inc 1 b2inc2 1 u when we estimate the
model. Another functional form misspecification occurs when we use the level of a variable when the
log of the variable is what actually shows up in the population model, or vice versa. For example, if
the true model has log(wage) as the dependent variable but we use wage as the dependent variable in
our regression analysis, then the estimators will be biased. Intuitively, this should be pretty clear. We
will discuss ways of detecting functional form misspecification in Chapter 9.
Omitting an important factor that is correlated with any of x1, x2, . . . , xk causes Assumption
MLR.4 to fail also. With multiple regression analysis, we are able to include many factors among
the explanatory variables, and omitted variables are less likely to be a problem in multiple regression
analysis than in simple regression analysis. Nevertheless, in any application, there are always factors
that, due to data limitations or ignorance, we will not be able to include. If we think these factors
should be controlled for and they are correlated with one or more of the independent variables, then
Assumption MLR.4 will be violated. We will derive this bias later.
There are other ways that u can be correlated with an explanatory variable. In Chapters 9
and 15, we will discuss the problem of measurement error in an explanatory variable. In Chapter 16,
we cover the conceptually more difficult problem in which one or more of the explanatory variables
is determined jointly with y—as occurs when we view quantities and prices as being determined by
the intersection of supply and demand curves. We must postpone our study of these problems until we
have a firm grasp of multiple regression analysis under an ideal set of assumptions.
When Assumption MLR.4 holds, we often say that we have exogenous explanatory variables.
If xj is correlated with u for any reason, then xj is said to be an endogenous explanatory variable.
58860_ch03_hr_066-116.indd 82
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
83
The terms “exogenous” and “endogenous” originated in simultaneous equations analysis (see
Chapter 16), but the term “endogenous explanatory variable” has evolved to cover any case in which
an explanatory variable may be correlated with the error term.
Before we show the unbiasedness of the OLS estimators under MLR.1 to MLR.4, a word of caution. Beginning students of econometrics sometimes confuse Assumptions MLR.3 and MLR.4, but
they are quite different. Assumption MLR.3 rules out certain relationships among the independent or
explanatory variables and has nothing to do with the error, u. You will know immediately when carrying out OLS estimation whether or not Assumption MLR.3 holds. On the other hand, Assumption
MLR.4—the much more important of the two—restricts the relationship between the unobserved
factors in u and the explanatory variables. Unfortunately, we will never know for sure whether the
average value of the unobserved factors is unrelated to the explanatory variables. But this is the critical assumption.
We are now ready to show unbiasedness of OLS under the first four multiple regression assumptions. As in the simple regression case, the expectations are conditional on the values of the explanatory variables in the sample, something we show explicitly in Appendix 3A but not in the text.
Theorem
3.1
Unbiasedness of OLS
Under Assumptions MLR.1 through MLR.4,
E 1 b^ j 2 5 bj, j 5 0, 1, . . . , k,
[3.37]
for any values of the population parameter bj. In other words, the OLS estimators are unbiased
estimators of the population parameters.
In our previous empirical examples, Assumption MLR.3 has been satisfied (because we have
been able to compute the OLS estimates). Furthermore, for the most part, the samples are randomly
chosen from a well-defined population. If we believe that the specified models are correct under the
key Assumption MLR.4, then we can conclude that OLS is unbiased in these examples.
Because we are approaching the point where we can use multiple regression in serious empirical
work, it is useful to remember the meaning of unbiasedness. It is tempting, in examples such as the
wage equation in (3.19), to say something like “9.2% is an unbiased estimate of the return to education.” As we know, an estimate cannot be unbiased: an estimate is a fixed number, obtained from a
particular sample, which usually is not equal to the population parameter. When we say that OLS is
unbiased under Assumptions MLR.1 through MLR.4, we mean that the procedure by which the OLS
estimates are obtained is unbiased when we view the procedure as being applied across all possible
random samples. We hope that we have obtained a sample that gives us an estimate close to the population value, but, unfortunately, this cannot be assured. What is assured is that we have no reason to
believe our estimate is more likely to be too big or more likely to be too small.
3-3a Including Irrelevant Variables in a Regression Model
One issue that we can dispense with fairly quickly is that of inclusion of an irrelevant variable or
overspecifying the model in multiple regression analysis. This means that one (or more) of the independent variables is included in the model even though it has no partial effect on y in the population.
(That is, its population coefficient is zero.)
To illustrate the issue, suppose we specify the model as
y 5 b0 1 b1x1 1 b2x2 1 b3x3 1 u,
[3.38]
and this model satisfies Assumptions MLR.1 through MLR.4. However, x3 has no effect on y after
x1 and x2 have been controlled for, which means that b3 5 0. The variable x3 may or may not be
58860_ch03_hr_066-116.indd 83
10/18/18 4:22 PM
84
PART 1 Regression Analysis with Cross-Sectional Data
correlated with x1 or x2; all that matters is that, once x1 and x2 are controlled for, x3 has no effect on y.
In terms of conditional expectations, E 1 y 0 x1, x2, x3 2 5 E 1 y 0 x1, x2 2 5 b0 1 b1x1 1 b2x2.
Because we do not know that b3 5 0, we are inclined to estimate the equation including x3:
y^ 5 b^ 0 1 b^ 1x1 1 b^ 2x2 1 b^ 3x3.
[3.39]
We have included the irrelevant variable, x3, in our regression. What is the effect of including x3 in
(3.39) when its coefficient in the population model (3.38) is zero? In terms of the unbiasedness of b^ 1
and b^ 2, there is no effect. This conclusion requires no special derivation, as it follows immediately from
Theorem 3.1. Remember, unbiasedness means E 1 b^ j 2 5 bj for any value of bj, including bj 5 0. Thus,
we can conclude that E 1 b^ 0 2 5 b0, E 1 b^ 1 2 5 b1, E 1 b^ 2 2 5 b2, E 1 b^ 3 2 5 0 (for any values of b0, b1,
and b2). Even though b^ 3 itself will never be exactly zero, its average value across all random samples
will be zero.
The conclusion of the preceding example is much more general: including one or more irrelevant
variables in a multiple regression model, or overspecifying the model, does not affect the unbiasedness of the OLS estimators. Does this mean it is harmless to include irrelevant variables? No. As we
will see in Section 3-4, including irrelevant variables can have undesirable effects on the variances of
the OLS estimators.
3-3b Omitted Variable Bias: The Simple Case
Now suppose that, rather than including an irrelevant variable, we omit a variable that actually
belongs in the true (or population) model. This is often called the problem of excluding a relevant
variable or underspecifying the model. We claimed in Chapter 2 and earlier in this chapter that this
problem generally causes the OLS estimators to be biased. It is time to show this explicitly and, just
as importantly, to derive the direction and size of the bias.
Deriving the bias caused by omitting an important variable is an example of misspecification
analysis. We begin with the case where the true population model has two explanatory variables and
an error term:
y 5 b0 1 b1x1 1 b2x2 1 u,
[3.40]
and we assume that this model satisfies Assumptions MLR.1 through MLR.4.
Suppose that our primary interest is in b1, the partial effect of x1 on y. For example, y is hourly
wage (or log of hourly wage), x1 is education, and x2 is a measure of innate ability. In order to get an
unbiased estimator of b1, we should run a regression of y on x1 and x2 (which gives unbiased estimators of b0, b1, and b2). However, due to our ignorance or data unavailability, we estimate the model by
excluding x2. In other words, we perform a simple regression of y on x1 only, obtaining the equation
| 1b
|x.
|
y5b
[3.41]
0
1 1
| comes from an underspecified model.
We use the symbol “|” rather than “^” to emphasize that b
1
When first learning about the omitted variable problem, it can be difficult to distinguish between
the underlying true model, (3.40) in this case, and the model that we actually estimate, which is captured by the regression in (3.41). It may seem silly to omit the variable x2 if it belongs in the model,
but often we have no choice. For example, suppose that wage is determined by
wage 5 b0 1 b1educ 1 b2abil 1 u.
[3.42]
Because ability is not observed, we instead estimate the model
wage 5 b0 1 b1educ 1 v,
where v 5 b2abil 1 u. The estimator of b1 from the simple regression of wage on educ is what we
|.
are calling b
1
58860_ch03_hr_066-116.indd 84
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
85
| conditional on the sample values of x and x . Deriving this
We derive the expected value of b
1
1
2
|
expectation is not difficult because b1 is just the OLS slope estimator from a simple regression, and
we have already studied this estimator extensively in Chapter 2. The difference here is that we must
analyze its properties when the simple regression model is misspecified due to an omitted variable.
As it turns out, we have done almost all of the work to derive the bias in the simple regression
| . From equation (3.23) we have the algebraic relationship b
| 5 b^ 1 b^ |
^
estimator of b
1
1
1
2 d 1, where b1
and b^ 2 are the slope estimators (if we could have them) from the multiple regression
yi on xi1, xi2, i 5 1, . . . , n
|
and d 1 is the slope from the simple regression
[3.43]
xi2 on xi1, i 5 1, . . . , n.
[3.44]
|
Because d 1 depends only on the independent variables in the sample, we treat it as fixed (nonrandom)
| ). Further, because the model in (3.40) satisfies Assumptions MLR.1 through
when computing E(b
1
MLR.4, we know that b^ 1 and b^ 2 would be unbiased for b1 and b2, respectively. Therefore,
| 2 5 E 1 b^ 1 b^ |
^
^ |
E1b
1
1
2 d 1 2 5 E 1 b1 2 1 E 1 b2 2 d 1
[3.45]
5 b1 1 b2 |
d 1,
| is
which implies the bias in b
1
| 2 5 E1b
| 2 2b 5b|
Bias 1 b
1
1
1
2 d 1.
[3.46]
Because the bias in this case arises from omitting the explanatory variable x2, the term on the righthand side of equation (3.46) is often called the omitted variable bias.
| is unbiased. The first is pretty
From equation (3.46), we see that there are two cases where b
1
| is unbiased. We
obvious: if b2 5 0—so that x2 does not appear in the true model (3.40)—then b
1
already know this from the simple regression analysis in Chapter 2. The second case is more interest| is unbiased for b , even if b 2 0.
ing. If |
d 1 5 0, then b
1
1
2
|
Because d 1 is the sample covariance between x1 and x2 over the sample variance of x1, |
d1 5 0
if, and only if, x1 and x2 are uncorrelated in the sample. Thus, we have the important conclusion that,
| is unbiased. This is not surprising: in Section 3-2,
if x1 and x2 are uncorrelated in the sample, then b
1
| and the multiple regression estimator b
| are the
we showed that the simple regression estimator b
1
1
|
same when x1 and x2 are uncorrelated in the sample. [We can also show that b1 is unbiased without
conditioning on the xi2 if E 1 x2 0 x1 2 5 E 1 x2 2 ; then, for estimating b1, leaving x2 in the error term does
not violate the zero conditional mean assumption for the error, once we adjust the intercept.]
When x1 and x2 are correlated, |
d 1 has the same sign as the correlation between x1 and x2: |
d1 . 0
if x1 and x2 are positively correlated and |
d 1 , 0& if x1 and x2 are negatively correlated. The sign of the
| depends on the signs of both b and d and is summarized in Table 3.2 for the four possible
bias in b
1
2
1
| is positive if
cases when there is bias. Table 3.2 warrants careful study. For example, the bias in b
1
b2 . 0 (x2 has a positive effect on y) and x1 and x2 are positively correlated, the bias is negative if
b2 . 0 and x1 and x2 are negatively correlated, and so on.
Table 3.2 summarizes the direction of the bias, but the size of the bias is also very important. A
small bias of either sign need not be a cause for concern. For example, if the return to education in the
population is 8.6% and the bias in the OLS estimator is 0.1% (a tenth of one percentage point), then
&
Table 3.2 Summary of Bias in b1 When x2 Is Omitted in
Estimating Equation (3.40)
58860_ch03_hr_066-116.indd 85
Corr 1 x1, x2 2 . 0
Corr 1 x1, x2 2 , 0
b2 . 0
Positive bias
Negative bias
b2 , 0
Negative bias
Positive bias
10/18/18 4:22 PM
86
PART 1 Regression Analysis with Cross-Sectional Data
we would not be very concerned. On the other hand, a bias on the order of three percentage points
would be much more serious. The size of the bias is determined by the sizes of b2 and |
d 1.
In practice, because b2 is an unknown population parameter, we cannot be certain whether b2 is
positive or negative. Nevertheless, we usually have a pretty good idea about the direction of the partial
effect of x2 on y. Further, even though the sign of the correlation between x1 and x2 cannot be known
if x2 is not observed, in many cases, we can make an educated guess about whether x1 and x2 are positively or negatively correlated.
In the wage equation (3.42), by definition, more ability leads to higher productivity and therefore
higher wages: b2 . 0. Also, there are reasons to believe that educ and abil are positively correlated:
on average, individuals with more innate ability choose higher levels of education. Thus, the OLS
estimates from the simple regression equation wage 5 b0 1 b1educ 1 v are on average too large.
This does not mean that the estimate obtained from our sample is too big. We can only say that if
we collect many random samples and obtain the simple regression estimates each time, then the average of these estimates will be greater than b1.
Example 3.6
Hourly Wage Equation
Suppose the model log 1 wage 2 5 b0 1 b1educ 1 b2abil 1 u satisfies Assumptions MLR.1 through
MLR.4. The data set in WAGE1 does not contain data on ability, so we estimate b1 from the simple
regression
log 1 wage 2 5 .584 1 .083 educ
n 5 526, R2 5 .186.
[3.47]
This is the result from only a single sample, so we cannot say that .083 is greater than b1; the true
return to education could be lower or higher than 8.3% (and we will never know for sure). Nevertheless,
we know that the average of the estimates across all random samples would be too large.
As a second example, suppose that, at the elementary school level, the average score for students
on a standardized exam is determined by
avgscore 5 b0 1 b1expend 1 b2povrate 1 u,
[3.48]
where expend is expenditure per student and povrate is the poverty rate of the children in the school.
Using school district data, we only have observations on the percentage of students with a passing
grade and per-student expenditures; we do not have information on poverty rates. Thus, we estimate
b1 from the simple regression of avgscore on expend.
| . First, b is probably negative: there is ample evidence
We can again obtain the likely bias in b
1
2
that children living in poverty score lower, on average, on standardized tests. Second, the average
expenditure per student is probably negatively correlated with the poverty rate: the higher the poverty
| will
rate, the lower the average per-student spending, so that Corr 1 x1, x2 2 , 0. From Table 3.2, b
1
have a positive bias. This observation has important implications. It could be that the true effect of
spending is zero; that is, b1 5 0. However, the simple regression estimate of b1 will usually be greater
than zero, and this could lead us to conclude that expenditures are important when they are not.
When reading and performing empirical work in economics, it is important to master the terminology associated with biased estimators. In the context of omitting a variable from model (3.40), if
| 2 . b , then we say that b
| has an upward bias. When E 1 b
| 2 ,b,b
| has a downward bias.
E1b
1
1
1
1
1
1
These definitions are the same whether b1 is positive or negative. The phrase biased toward zero
| 2 is closer to zero than is b . Therefore, if b is positive, then b
| is biased
refers to cases where E 1 b
1
1
1
1
|
toward zero if it has a downward bias. On the other hand, if b1 , 0, then b1 is biased toward zero if it
has an upward bias.
58860_ch03_hr_066-116.indd 86
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
87
3-3c Omitted Variable Bias: More General Cases
Deriving the sign of omitted variable bias when there are multiple regressors in the estimated model is
more difficult. We must remember that correlation between a single explanatory variable and the error
generally results in all OLS estimators being biased. For example, suppose the population model
y 5 b0 1 b1x1 1 b2x2 1 b3x3 1 u
satisfies Assumptions MLR.1 through MLR.4. But we omit x3 and estimate the model as
| 1b
|x 1b
|x.
|
y5b
0
1 1
2 2
[3.49]
[3.50]
Now, suppose that x2 and x3 are uncorrelated, but that x1 is correlated with x3. In other words, x1 is
| is probably
correlated with the omitted variable, but x2 is not. It is tempting to think that, while b
1
|
biased based on the derivation in the previous subsection, b2 is unbiased because x2 is uncorrelated
| and b
| will normally be biased. The only
with x3. Unfortunately, this is not generally the case: both b
1
2
exception to this is when x1 and x2 are also uncorrelated.
| and
Even in the fairly simple model above, it can be difficult to obtain the direction of bias in b
1
|
b2. This is because x1, x2, and x3 can all be pairwise correlated. Nevertheless, an approximation is
|
often practically useful. If we assume that x1 and x2 are uncorrelated, then we can study the bias in b
1
as if x2 were absent from both the population and the estimated models. In fact, when x1 and x2 are
uncorrelated, it can be shown that
a 1 xi1 2 x1 2 xi3
n
| 2 5b 1b
E1b
1
1
3
i51
n
2
a 1 xi1 2 x1 2
.
i51
This is just like equation (3.45), but b3 replaces b2, and x3 replaces x2 in regression (3.44). Therefore,
| is obtained by replacing b with b and x with x in Table 3.2. If b . 0 and
the bias in b
1
2
3
2
3
3
| is positive, and so on.
Corr 1 x1, x3 2 . 0, the bias in b
1
As an example, suppose we add exper to the wage model:
wage 5 b0 1 b1educ 1 b2exper 1 b3abil 1 u.
If abil is omitted from the model, the estimators of both b1 and b2 are biased, even if we assume exper
is uncorrelated with abil. We are mostly interested in the return to education, so it would be nice if
| has an upward or a downward bias due to omitted ability. This conclusion
we could conclude that b
1
is not possible without further assumptions. As an approximation, let us suppose that, in addition to
exper and abil being uncorrelated, educ and exper are also uncorrelated. (In reality, they are some| would
what negatively correlated.) Because b3 . 0 and educ and abil are positively correlated, b
1
have an upward bias, just as if exper were not in the model.
The reasoning used in the previous example is often followed as a rough guide for obtaining
the likely bias in estimators in more complicated models. Usually, the focus is on the relationship
between a particular explanatory variable, say, x1, and the key omitted factor. Strictly speaking, ignoring all other explanatory variables is a valid practice only when each one is uncorrelated with x1, but
it is still a useful guide. Appendix 3A contains a more careful analysis of omitted variable bias with
multiple explanatory variables.
3-4 The Variance of the OLS Estimators
We now obtain the variance of the OLS estimators so that, in addition to knowing the central tendencies of the b^ j, we also have a measure of the spread in its sampling distribution. Before finding
the variances, we add a homoskedasticity assumption, as in Chapter 2. We do this for two reasons.
58860_ch03_hr_066-116.indd 87
10/18/18 4:22 PM
88
PART 1 Regression Analysis with Cross-Sectional Data
First, the formulas are simplified by imposing the constant error variance assumption. Second, in
Section 3-5, we will see that OLS has an important efficiency property if we add the homoskedasticity
assumption.
In the multiple regression framework, homoskedasticity is stated as follows:
Assumption MLR.5
Homoskedasticity
The error u has the same variance given any value of the explanatory variables. In other words,
Var 1 u 0 x1, . . . , x k 2 5 s2.
Assumption MLR.5 means that the variance in the error term, u, conditional on the explanatory variables, is the same for all combinations of outcomes of the explanatory variables. If this assumption
fails, then the model exhibits heteroskedasticity, just as in the two-variable case.
In the equation
wage 5 b0 1 b1educ 1 b2exper 1 b3tenure 1 u,
homoskedasticity requires that the variance of the unobserved error u does not depend on the levels of
education, experience, or tenure. That is,
Var 1 u 0 educ, exper, tenure 2 5 s2.
If this variance changes with any of the three explanatory variables, then heteroskedasticity is
present.
Assumptions MLR.1 through MLR.5 are collectively known as the Gauss-Markov assumptions
(for cross-sectional regression). So far, our statements of the assumptions are suitable only when
applied to cross-sectional analysis with random sampling. As we will see, the Gauss-Markov assumptions for time series analysis, and for other situations such as panel data analysis, are more difficult to
state, although there are many similarities.
In the discussion that follows, we will use the symbol x to denote the set of all independent
variables, 1 x1, . . . , xk 2 . Thus, in the wage regression with educ, exper, and tenure as independent variables, x 5 1 educ, exper, tenure 2 . Then we can write Assumptions MLR.1 and MLR.4 as
E 1 y 0 x 2 5 b0 1 b1x1 1 b2x2 1 p 1 bkxk,
and Assumption MLR.5 is the same as Var 1 y 0 x 2 5 s2. Stating the assumptions in this way clearly
illustrates how Assumption MLR.5 differs greatly from Assumption MLR.4. Assumption MLR.4
says that the expected value of y, given x, is linear in the parameters, but it certainly depends on
x1, x2, c, xk. Assumption MLR.5 says that the variance of y, given x, does not depend on the values
of the independent variables.
We can now obtain the variances of the b^ j, where we again condition on the sample values of the
independent variables. The proof is in the appendix to this chapter.
Theorem
3.2
Sampling Variances of the OLS Slope Estimators
Under Assumptions MLR.1 through MLR.5, conditional on the sample values of the independent
variables,
Var 1 b^ j 2 5
s2
SSTj 1 1 2 R 2j 2 r
[3.51]
for j 5 1, 2, . . . , k , where SSTj 5 a ni51 1 xij 2 xj 2 2 is the total sample variation in xj , and R2j is the
R-squared from regressing xj on all other independent variables (and including an intercept).
58860_ch03_hr_066-116.indd 88
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
89
The careful reader may be wondering whether there is a simple formula for the variance of b^ j
where we do not condition on the sample outcomes of the explanatory variables. The answer is: none
that is useful. The formula in (3.51) is a highly nonlinear function of the xij, making averaging out
across the population distribution of the explanatory variables virtually impossible. Fortunately, for
any practical purpose equation (3.51) is what we want. Even when we turn to approximate, largesample properties of OLS in Chapter 5 it turns out that (3.51) estimates the quantity we need for
large-sample analysis, provided Assumptions MLR.1 through MLR.5 hold.
Before we study equation (3.51) in more detail, it is important to know that all of the GaussMarkov assumptions are used in obtaining this formula. Whereas we did not need the homoskedasticity assumption to conclude that OLS is unbiased, we do need it to justify equation (3.51).
The size of Var 1 b^ j 2 is practically important. A larger variance means a less precise estimator, and
this translates into larger confidence intervals and less accurate hypotheses tests (as we will see in
Chapter 4). In the next subsection, we discuss the elements comprising (3.51).
3-4a The Components of the OLS Variances: Multicollinearity
Equation (3.51) shows that the variance of b^ j depends on three factors: s2, SSTj, and R2j . Remember
that the index j simply denotes any one of the independent variables (such as education 3 or poverty
rate). We now consider each of the factors affecting Var 1 b^ j 2 in turn.
The Error Variance, s2. From equation (3.51), a larger s2 means larger sampling variances
for the OLS estimators. This is not at all surprising: more “noise” in the equation (a larger s2) makes it
more difficult to estimate the partial effect of any of the independent variables on y, and this is reflected
in higher variances for the OLS slope estimators. Because s2 is a feature of the population, it has
nothing to do with the sample size. It is the one component of (3.51) that is unknown. We will see later
how to obtain an unbiased estimator of s2.
For a given dependent variable y, there is really only one way to reduce the error variance, and
that is to add more explanatory variables to the equation (take some factors out of the error term).
Unfortunately, it is not always possible to find additional legitimate factors that affect y.
The Total Sample Variation in xj , SSTj . From equation (3.51), we see that the larger
the total variation in xj is, the smaller is Var 1 b^ j 2 . Thus, everything else being equal, for estimating bj
we prefer to have as much sample variation in xj as possible. We already discovered this in the simple
regression case in Chapter 2. Although it is rarely possible for us to choose the sample values of the
independent variables, there is a way to increase the sample variation in each of the independent
variables: increase the sample size. In fact, when one randomly samples from a population, SSTj
increases without bound as the sample size increases—roughly as a linear function of n. This is the
component of the variance that systematically depends on the sample size.
When SSTj is small, Var 1 b^ j 2 can get very large, but a small SSTj is not a violation of Assumption
MLR.3. Technically, as SSTj goes to zero, Var 1 b^ j 2 approaches infinity. The extreme case of no sample variation in xj, SSTj 5 0, is not allowed by Assumption MLR.3 because then we cannot even
compute the OLS estimates.
The Linear Relationships among the Independent Variables, Rj2. The
term R2j in equation (3.51) is the most difficult of the three components to understand. This term
does not appear in simple regression analysis because there is only one independent variable in such
cases. It is important to see that this R-squared is distinct from the R‑squared in the regression of y on
x1, x2, . . . , xk: R2j is obtained from a regression involving only the independent variables in the original
model, where xj plays the role of a dependent variable.
Consider first the k 5 2 case: y 5 b0 1 b1x1 1 b2x2 1 u. Then, Var 1 b^ 1 2 5 s2 / 3 SST1 1 1 2 R21 2 4 ,
where R21 is the R-squared from the simple regression of x1 on x2 (and an intercept, as always).
58860_ch03_hr_066-116.indd 89
10/18/18 4:22 PM
90
PART 1 Regression Analysis with Cross-Sectional Data
Figure 3.1 Var 1 b^ 1 2 as a function of R 12.
Var( ˆ 1)
0
R12
1
Because the R-squared measures goodness-of-fit, a value of R21 close to one indicates that x2 explains
much of the variation in x1 in the sample. This means that x1 and x2 are highly correlated.
As R21 increases to one, Var 1 b^ 1 2 gets larger and larger. Thus, a high degree of linear relationship
between x1 and x2 can lead to large variances for the OLS slope estimators. (A similar argument applies to
b^ 2.) See Figure 3.1 for the relationship between Var 1 b^ 1 2 and the R-squared from the regression of x1 on x2.
In the general case, R2j is the proportion of the total variation in xj that can be explained by the
other independent variables appearing in the equation. For a given s2 and SSTj, the smallest Var 1 b^ j 2
is obtained when R2j 5 0, which happens if, and only if, xj has zero sample correlation with every
other independent variable. This is the best case for estimating bj, but it is rarely encountered.
The other extreme case, R2j 5 1, is ruled out by Assumption MLR.3, because R2j 5 1 means that,
in the sample, xj is a perfect linear combination of some of the other independent variables in the
regression. A more relevant case is when R2j is “close” to one. From equation (3.51) and Figure 3.1,
we see that this can cause Var 1 b^ j 2 to be large: Var 1 b^ j 2 S ` as R2j S 1. High (but not perfect) correlation between two or more independent variables is called multicollinearity.
Before we discuss the multicollinearity issue further, it is important to be very clear on one thing:
a case where R2j is close to one is not a violation of Assumption MLR.3.
Because multicollinearity violates none of our assumptions, the “problem” of multicollinearity is not
really well defined. When we say that multicollinearity arises for estimating bj when R2j is “close” to one,
we put “close” in quotation marks because there is no absolute number that we can cite to conclude that
multicollinearity is a problem. For example, R2j 5 .9 means that 90% of the sample variation in xj can be
explained by the other independent variables in the regression model. Unquestionably, this means that
xj has a strong linear relationship to the other independent variables. But whether this translates into a
Var 1 b^ j 2 that is too large to be useful depends on the sizes of s2 and SSTj. As we will see in Chapter 4, for
statistical inference, what ultimately matters is how big b^ j is in relation to its standard deviation.
58860_ch03_hr_066-116.indd 90
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
91
Just as a large value of R2j can cause a large Var 1 b^ j 2 , so can a small value of SSTj. Therefore,
a small sample size can lead to large sampling variances, too. Worrying about high degrees of correlation among the independent variables in the sample is really no different from worrying about a
small sample size: both work to increase Var 1 b^ j 2 . The famous University of Wisconsin econometrician Arthur Goldberger, reacting to econometricians’ obsession with multicollinearity, has (tongue in
cheek) coined the term micronumerosity, which he defines as the “problem of small sample size.”
[For an engaging discussion of multicollinearity and micronumerosity, see Goldberger (1991).]
Although the problem of multicollinearity cannot be clearly defined, one thing is clear: everything else being equal, for estimating bj, it is better to have less correlation between xj and the other
independent variables. This observation often leads to a discussion of how to “solve” the multicollinearity problem. In the social sciences, where we are usually passive collectors of data, there is no
good way to reduce variances of unbiased estimators other than to collect more data. For a given data
set, we can try dropping other independent variables from the model in an effort to reduce multicollinearity. Unfortunately, dropping a variable that belongs in the population model can lead to bias, as
we saw in Section 3-3.
Perhaps an example at this point will help clarify some of the issues raised concerning multicollinearity. Suppose we are interested in estimating the effect of various school expenditure categories
on student performance. It is likely that expenditures on teacher salaries, instructional materials, athletics, and so on are highly correlated: wealthier schools tend to spend more on everything, and poorer
schools spend less on everything. Not surprisingly, it can be difficult to estimate the effect of any
particular expenditure category on student performance when there is little variation in one category
that cannot largely be explained by variations in the other expenditure categories (this leads to high R2j
for each of the expenditure variables). Such multicollinearity problems can be mitigated by collecting
more data, but in a sense we have imposed the problem on ourselves: we are asking questions that
may be too subtle for the available data to answer with any precision. We can probably do much better by changing the scope of the analysis and lumping all expenditure categories together, because we
would no longer be trying to estimate the partial effect of each separate category.
Another important point is that a high degree of correlation between certain independent variables can be irrelevant as to how well we can estimate other parameters in the model. For example,
consider a model with three independent variables:
y 5 b0 1 b1x1 1 b2x2 1 b3x3 1 u,
where x2 and x3 are highly correlated. Then Var 1 b^ 2 2 and Var 1 b^ 3 2 may be large. But the amount
of correlation between x2 and x3 has no direct effect on Var 1 b^ 1 2 . In fact, if x1 is uncorrelated with x2
and x3, then R21 5 0 and Var 1 b^ 1 2 5 s2 /SST1, regardless of how much correlation there is between x2 and x3.
G o i n g F u rt h e r 3 . 4
If b1 is the parameter of interest, we do not really care
Suppose you postulate a model explaining
about the amount of correlation between x2 and x3.
final exam score in terms of class attenThe previous observation is important because
dance. Thus, the dependent variable is
economists often include many control variables
final exam score, and the key explanatory
in order to isolate the causal effect of a particular
variable is number of classes attended. To
variable. For example, in looking at the relationcontrol for student abilities and efforts outside the classroom, you include among the
ship between loan approval rates and percentage
explanatory variables cumulative GPA, SAT
of minorities in a neighborhood, we might include
score, and measures of high school perforvariables like average income, average housing
mance. Someone says, “You cannot hope
value, measures of creditworthiness, and so on,
to learn anything from this exercise because
because these factors need to be accounted for in
cumulative GPA, SAT score, and high school
order to draw causal conclusions about discriminaperformance are likely to be highly collinear.”
tion. Income, housing prices, and creditworthiness
What should be your response?
are generally highly correlated with each other.
58860_ch03_hr_066-116.indd 91
10/18/18 4:22 PM
92
PART 1 Regression Analysis with Cross-Sectional Data
But high correlations among these controls do not make it more difficult to determine the effects
of discrimination.
Some researchers find it useful to compute statistics intended to determine the severity of multicollinearity in a given application. Unfortunately, it is easy to misuse such statistics because, as we
have discussed, we cannot specify how much correlation among explanatory variables is “too much.”
Some multicollinearity “diagnostics” are omnibus statistics in the sense that they detect a strong linear
relationship among any subset of explanatory variables. For reasons that we just saw, such statistics
are of questionable value because they might reveal a “problem” simply because two control variables,
whose coefficients we do not care about, are highly correlated. [Probably the most common omnibus
multicollinearity statistic is the so-called condition number, which is defined in terms of the full data
matrix and is beyond the scope of this text. See, for example, Belsley, Kuh, and Welsh (1980).]
Somewhat more useful, but still prone to misuse, are statistics for individual coefficients. The
most common of these is the variance inflation factor (VIF), which is obtained directly from equation (3.51). The VIF for slope coefficient j is simply VIFj 5 1/ 1 1 2 R2j 2 , precisely the term in Var 1 b^ j 2
that is determined by correlation between xj and the other explanatory variables. We can write Var 1 b^ j 2
in equation (3.51) as
Var 1 b^ j 2 5
s2
? VIFj,
SSTj
which shows that VIFj is the factor by which Var 1 b^ j 2 is higher because xj is not uncorrelated with
the other explanatory variables. Because VIFj is a function of R2j —indeed, Figure 3.1 is essentially a
graph of VIF1—our previous discussion can be cast entirely in terms of the VIF. For example, if we
had the choice, we would like VIFj to be smaller (other things equal). But we rarely have the choice.
if we think certain explanatory variables need to be included in a regression to infer causality of xj,
then we are hesitant to drop them, and whether we think VIFj is “too high” cannot really affect that
decision. If, say, our main interest is in the causal effect of x1 on y, then we should ignore entirely the
VIFs of other coefficients. Finally, setting a cutoff value for VIF above which we conclude multicollinearity is a “problem” is arbitrary and not especially helpful. Sometimes the value 10 is chosen: if
VIFj is above 10 (equivalently, R2j is above .9), then we conclude that multicollinearity is a “problem”
for estimating bj. But a VIFj above 10 does not mean that the standard deviation of b^ j is too large to be
useful because the standard deviation also depends on s and SSTj, and the latter can be increased by
increasing the sample size. Therefore, just as with looking at the size of R2j directly, looking at the size
of VIFj is of limited use, although one might want to do so out of curiosity.
3-4b Variances in Misspecified Models
The choice of whether to include a particular variable in a regression model can be made by analyzing
the tradeoff between bias and variance. In Section 3-3, we derived the bias induced by leaving out a
relevant variable when the true model contains two explanatory variables. We continue the analysis of
this model by comparing the variances of the OLS estimators.
Write the true population model, which satisfies the Gauss-Markov assumptions, as
y 5 b0 1 b1x1 1 b2x2 1 u.
We consider two estimators of b1. The estimator b^ 1 comes from the multiple regression
y^ 5 b^ 0 1 b^ 1x1 1 b^ 2x2.
[3.52]
| is obtained by
In other words, we include x2, along with x1, in the regression model. The estimator b
1
omitting x2 from the model and running a simple regression of y on x1:
| 1b
|x.
|
y5b
[3.53]
0
58860_ch03_hr_066-116.indd 92
1 1
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
93
When b2 2 0, equation (3.53) excludes a relevant variable from the model and, as we saw in
| unless x and x are uncorrelated. On the other hand, b^ is unbiSection 3-3, this induces a bias in b
1
1
2
1
ased for b1 for any value of b2, including b2 5 0. It follows that, if bias is used as the only criterion,
|.
b^ 1 is preferred to b
1
| does not carry over when we bring variance into
The conclusion that b^ 1 is always preferred to b
1
the picture. Conditioning on the values of x1 and x2 in the sample, we have, from (3.51),
Var 1 b^ 1 2 5 s2 / 3 SST1 1 1 2 R21 2 4 ,
[3.54]
R21
where SST1 is the total variation in x1, and is the R-squared from the regression of x1 on x2. Further,
a simple modification of the proof in Chapter 2 for two-variable regression shows that
| 2 5 s2 /SST .
Var 1 b
[3.55]
1
1
| 2 is always smaller than Var 1 b^ 2 , unless x and x are
Comparing (3.55) to (3.54) shows that Var 1 b
1
1
1
2
| and b^ are the same. Assuming that x
uncorrelated in the sample, in which case the two estimators b
1
1
1
and x2 are not uncorrelated, we can draw the following conclusions:
| is biased, b^ is unbiased, and Var 1 b
| 2 , Var 1 b^ 2 .
1. When b 2 0, b
2
1
1
1
1
| and b^ are both unbiased, and Var 1 b
| 2 , Var 1 b^ 2 .
2. When b2 5 0, b
1
1
1
1
|
From the second conclusion, it is clear that b1 is preferred if b2 5 0. Intuitively, if x2 does not have a
partial effect on y, then including it in the model can only exacerbate the multicollinearity problem,
which leads to a less efficient estimator of b1. A higher variance for the estimator of b1 is the cost of
including an irrelevant variable in a model.
The case where b2 2 0 is more difficult. Leaving x2 out of the model results in a biased estimator
of b1. Traditionally, econometricians have suggested comparing the likely size of the bias due to omitting x2 with the reduction in the variance—summarized in the size of R21—to decide whether x2 should
be included. However, when b2 2 0, there are two favorable reasons for including x2 in the model.
| does not shrink as the sample size grows; in fact,
The most important of these is that any bias in b
1
the bias does not necessarily follow any pattern. Therefore, we can usefully think of the bias as being
| 2 and Var 1 b^ 2 both shrink to zero as
roughly the same for any sample size. On the other hand, Var 1 b
1
1
n gets large, which means that the multicollinearity induced by adding x2 becomes less important as
the sample size grows. In large samples, we would prefer b^ 1.
The other reason for favoring b^ 1 is more subtle. The variance formula in (3.55) is conditional on
| . When b 2 0, the varithe values of xi1 and xi2 in the sample, which provides the best scenario for b
1
2
|
ance of b1 conditional only on x1 is larger than that presented in (3.55). Intuitively, when b2 2 0 and
x2 is excluded from the model, the error variance increases because the error effectively contains part
of x2. But the expression in equation (3.55) ignores the increase in the error variance because it will
treat both regressors as nonrandom. For practical purposes, the s2 term in equation (3.55) increases
when x2 is dropped from the equation. A full discussion of the proper conditioning argument when
computing the OLS variances would lead us too far astray. Suffice it to say that equation (3.55) is too
| . Fortunately, statistical packages report the
generous when it comes to measuring the precision of b
1
proper variance estimator, and so we need not worry about the subtleties in the theoretical formulas.
After reading the next subsection, you might want to study Problems 14 and 15 for further insight.
3-4c Estimating s2: Standard Errors of the OLS Estimators
We now show how to choose an unbiased estimator of s2, which then allows us to obtain unbiased
estimators of Var 1 b^ j 2 .
Because s2 5 E 1 u2 2 , an unbiased “estimator” of s2 is the sample average of the squared errors:
21 n
n g i51 u2i . Unfortunately, this is not a true estimator because we do not observe the ui. Nevertheless,
recall that the errors can be written as ui 5 yi 2 b0 2 b1xi1 2 b2xi2 2 p 2 bkxik, and so the reason
58860_ch03_hr_066-116.indd 93
10/18/18 4:22 PM
94
PART 1 Regression Analysis with Cross-Sectional Data
we do not observe the ui is that we do not know the bj. When we replace each bj with its OLS estimator, we get the OLS residuals:
u^ i 5 yi 2 b^ 0 2 b^ 1xi1 2 b^ 2xi2 2 p 2 b^ kxik.
It seems natural to estimate s2 by replacing ui with the u^ i. In the simple regression case, we saw
that this leads to a biased estimator. The unbiased estimator of s2 in the general multiple regression
case is
s^ 2 5 a a u^ 2i b ^ 1 n 2 k 2 1 2 5 SSR/ 1 n 2 k 2 1 2 .
n
i51
[3.56]
We already encountered this estimator in the k 5 1 case in simple regression.
The term n − k − 1 in (3.56) is the degrees of freedom (df) for the general OLS problem with n
observations and k independent variables. Because there are k + 1 parameters in a regression model
with k independent variables and an intercept, we can write
df 5 n 2 1 k 1 1 2
5 1 number of observations 2 2 1 number of estimated parameters 2 .
[3.57]
This is the easiest way to compute the degrees of freedom in a particular application: count the number of parameters, including the intercept, and subtract this amount from the number of observations.
(In the rare case that an intercept is not estimated, the number of parameters decreases by one.)
Technically, the division by n − k − 1 in (3.56) comes from the fact that the expected value of the
sum of squared residuals is E 1 SSR 2 5 1 n 2 k 2 1 2 s2. Intuitively, we can figure out why the degrees
of freedom adjustment is necessary by returning to the first order conditions for the OLS estimators.
n
n
These can be written g i51 u^ i 5 0 and g i51xiju^ i 5 0, where j 5 1, 2, p , k. Thus, in obtaining the
OLS estimates, k 1 1 restrictions are imposed on the OLS residuals. This means that, given n − (k 1 1)
of the residuals, the remaining k 1 1 residuals are known: there are only n 2 (k 1 1) degrees of freedom
in the residuals. (This can be contrasted with the errors ui, which have n degrees of freedom in the sample.)
For reference, we summarize this discussion with Theorem 3.3. We proved this theorem for the
case of simple regression analysis in Chapter 2 (see Theorem 2.3). (A general proof that requires
matrix algebra is provided in Advanced Treatment E.)
Theorem
3.3
Unbiased Estimation of s2
Under the Gauss-Markov assumptions MLR.1 through MLR.5, E 1 s^ 2 2 5 s2.
The positive square root of s^ 2, denoted s^ , is called the standard error of the regression (SER). The
SER is an estimator of the standard deviation of the error term. This estimate is usually reported by
regression packages, although it is called different things by different packages. (In addition to SER,
s^ is also called the standard error of the estimate and the root mean squared error.)
Note that s^ can either decrease or increase when another independent variable is added to a
regression (for a given sample). This is because, although SSR must fall when another explanatory
variable is added, the degrees of freedom also falls by one. Because SSR is in the numerator and df is
in the denominator, we cannot tell beforehand which effect will dominate.
For constructing confidence intervals and conducting tests in Chapter 4, we will need to estimate
the standard deviation of b^ j, which is just the square root of the variance:
sd 1 b^ j 2 5 s/ 3 SSTj 1 1 2 R2j 2 4 1/2.
Because s is unknown, we replace it with its estimator, s^ . This gives us the standard error of b^ j:
se 1 b^ j 2 5 s^ / 3 SSTj 1 1 2 R2j 2 4 1/2.
58860_ch03_hr_066-116.indd 94
[3.58]
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
95
Just as the OLS estimates can be obtained for any given sample, so can the standard errors. Because
se 1 b^ j 2 depends on s^ , the standard error has a sampling distribution, which will play a role in Chapter 4.
We should emphasize one thing about standard errors. Because (3.58) is obtained directly from
the variance formula in (3.51), and because (3.51) relies on the homoskedasticity Assumption MLR.5,
it follows that the standard error formula in (3.58) is not a valid estimator of sd 1 b^ j 2 if the errors
exhibit heteroskedasticity. Thus, while the presence of heteroskedasticity does not cause bias in the b^ j,
it does lead to bias in the usual formula for Var 1 b^ j 2 , which then invalidates the standard errors. This is
important because any regression package computes (3.58) as the default standard error for each coefficient (with a somewhat different representation for the intercept). If we suspect heteroskedasticity,
then the “usual” OLS standard errors are invalid, and some corrective action should be taken. We will
see in Chapter 8 what methods are available for dealing with heteroskedasticity.
For some purposes it is helpful to write
se 1 b^ j 2 5
s^
!nsd 1 xj 2 "1 2 R2j
,
[3.59]
n
in which we take sd 1 xj 2 5 "n 21 g i51 1 xij 2 xj 2 2 to be the sample standard deviation where the total
sum of squares is divided by n rather than n − 1. The importance of equation (3.59) is that it shows
how the sample size, n, directly affects the standard errors. The other three terms in the formula—s^ ,
sd 1 xj 2 , and R2j —will change with different samples, but as n gets large they settle down to constants.
Therefore, we can see from equation (3.59) that the standard errors shrink to zero at the rate 1/ !n.
This formula demonstrates the value of getting more data: the precision of the b^ j increases as n
increases. (By contrast, recall that unbiasedness holds for any sample size subject to being able to
compute the estimators.) We will talk more about large sample properties of OLS in Chapter 5.
3-5 Efficiency of OLS: The Gauss-Markov Theorem
In this section, we state and discuss the important Gauss-Markov Theorem, which justifies the use
of the OLS method rather than using a variety of competing estimators. We know one justification for
OLS already: under Assumptions MLR.1 through MLR.4, OLS is unbiased. However, there are many
unbiased estimators of the bj under these assumptions (for example, see Problem 13). Might there be
other unbiased estimators with variances smaller than the OLS estimators?
If we limit the class of competing estimators appropriately, then we can show that OLS is best
within this class. Specifically, we will argue that, under Assumptions MLR.1 through MLR.5, the
OLS estimator b^ j for bj is the best linear unbiased estimator (BLUE). To state the theorem, we
need to understand each component of the acronym “BLUE.” First, we know what an estimator is:
it is a rule that can be applied to any sample of data to produce an estimate. We also know what an
| , of b is an unbiased estimator of b
unbiased estimator is: in the current context, an estimator, say, b
1
j
j
|
if E 1 bj 2 5 bj for any b0, b1, . . . , bk.
| of b is linWhat about the meaning of the term “linear”? In the current context, an estimator b
j
j
ear if, and only if, it can be expressed as a linear function of the data on the dependent variable:
| 5
b
a wijyi,
j
n
[3.60]
i51
where each wij can be a function of the sample values of all the independent variables. The OLS estimators are linear, as can be seen from equation (3.22).
Finally, how do we define “best”? For the current theorem, best is defined as having the smallest
variance. Given two unbiased estimators, it is logical to prefer the one with the smallest variance (see
Math Refresher C).
58860_ch03_hr_066-116.indd 95
10/18/18 4:22 PM
96
PART 1 Regression Analysis with Cross-Sectional Data
^ 1, c, b
^ k, denote the OLS estimators in model (3.31) under Assumptions MLR.1
Now, let b^ 0, b
| that is linear and unbi‑
through MLR.5. The Gauss-Markov Theorem says that, for any estimator b
j
|
ased, Var 1 b^ j 2 # Var 1 bj 2 , and the inequality is usually strict. In other words, in the class of linear
unbiased estimators, OLS has the smallest variance (under the five Gauss-Markov assumptions).
Actually, the theorem says more than this. If we want to estimate any linear function of the bj, then
the corresponding linear combination of the OLS estimators achieves the smallest variance among
all linear unbiased estimators. We conclude with a theorem, which is proven in Appendix 3A. It is
because of this theorem that Assumptions MLR.1 through MLR.5 are known as the Gauss-Markov
assumptions (for cross-sectional analysis).
Theorem
3.4
Gauss-Markov Theorem
Under Assumptions MLR.1 through MLR.5, b^ 0, b^ 1, c, b^ k are the best linear unbiased estimators
(BLUEs) of b0, b1, . . . , bk, respectively.
The importance of the Gauss-Markov Theorem is that, when the standard set of assumptions
holds, we need not look for alternative unbiased estimators of the form in (3.60): none will be better
than OLS. Equivalently, if we are presented with an estimator that is both linear and unbiased, then
we know that the variance of this estimator is at least as large as the OLS variance; no additional calculation is needed to show this.
For our purposes, Theorem 3.4 justifies the use of OLS to estimate multiple regression models. If
any of the Gauss-Markov assumptions fail, then this theorem no longer holds. We already know that
failure of the zero conditional mean assumption (Assumption MLR.4) causes OLS to be biased, so
Theorem 3.4 also fails. We also know that heteroskedasticity (failure of Assumption MLR.5) does not
cause OLS to be biased. However, OLS no longer has the smallest variance among linear unbiased
estimators in the presence of heteroskedasticity. In Chapter 8, we analyze an estimator that improves
upon OLS when we know the brand of heteroskedasticity.
3-6 Some Comments on the Language of Multiple
Regression Analysis
It is common for beginners, and not unheard of for experienced empirical researchers, to report that
they “estimated an OLS model.” Although we can usually figure out what someone means by this
statement, it is important to understand that it is wrong—on more than just an aesthetic level—and
reflects a misunderstanding about the components of a multiple regression analysis.
The first thing to remember is that ordinary least squares (OLS) is an estimation method, not a
model. A model describes an underlying population and depends on unknown parameters. The linear
model that we have been studying in this chapter can be written—in the population—as
y 5 b0 1 b1x1 1 p 1 bkxk 1 u,
[3.61]
where the parameters are the bj. Importantly, we can talk about the meaning of the bj without ever
looking at data. It is true we cannot hope to learn much about the bj without data, but the interpretation of the bj is obtained from the linear model in equation (3.61).
Once we have a sample of data we can estimate the parameters. Although it is true that we have
so far only discussed OLS as a possibility, there are actually many more ways to use the data than we
can even list. We have focused on OLS due to its widespread use, which is justified by using the statistical considerations we covered previously in this chapter. But the various justifications for OLS rely
on the assumptions we have made (MLR.1 through MLR.5). As we will see in later chapters, under
58860_ch03_hr_066-116.indd 96
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
97
different assumptions different estimation methods are preferred—even though our model can still be
represented by equation (3.61). Just a few examples include weighted least squares in Chapter 8, least
absolute deviations in Chapter 9, and instrumental variables in Chapter 15.
One might argue that the discussion here is overly pedantic, and that the phrase “estimating an
OLS model” should be taken as a useful shorthand for “I estimated a linear model by OLS.” This
stance has some merit, but we must remember that we have studied the properties of the OLS estimators under different assumptions. For example, we know OLS is unbiased under the first four GaussMarkov assumptions, but it has no special efficiency properties without Assumption MLR.5. We have
also seen, through the study of the omitted variables problem, that OLS is biased if we do not have
Assumption MLR.4. The problem with using imprecise language is that it leads to vagueness on
the most important considerations: what assumptions are being made on the underlying linear model?
The issue of the assumptions we are using is conceptually different from the estimator we wind up
applying.
Ideally, one writes down an equation like (3.61), with variable names that are easy to decipher,
such as
math4 5 b0 1 b1classize4 1 b2math3 1 b3log 1 income 2
1 b4motheduc 1 b5 fatheduc 1 u
[3.62]
if we are trying to explain outcomes on a fourth-grade math test. Then, in the context of equation
(3.62), one includes a discussion of whether it is reasonable to maintain Assumption MLR.4, focusing on the factors that might still be in u and whether more complicated functional relationships are
needed (a topic we study in detail in Chapter 6). Next, one describes the data source (which ideally
is obtained via random sampling) as well as the OLS estimates obtained from the sample. A proper
way to introduce a discussion of the estimates is to say “I estimated equation (3.62) by ordinary least
squares. Under the assumption that no important variables have been omitted from the equation, and
assuming random sampling, the OLS estimator of the class size effect, b1, is unbiased. If the error
term u has constant variance, the OLS estimator is actually best linear unbiased.” As we will see in
Chapters 4 and 5, we can often say even more about OLS. Of course, one might want to admit that
while controlling for third-grade math score, family income and parents’ education might account for
important differences across students, it might not be enough—for example, u can include motivation
of the student or parents—in which case OLS might be biased.
A more subtle reason for being careful in distinguishing between an underlying population model
and an estimation method used to estimate a model is that estimation methods such as OLS can be
used essentially as an exercise in curve fitting or prediction, without explicitly worrying about an
underlying model and the usual statistical properties of unbiasedness and efficiency. For example, we
might just want to use OLS to estimate a line that allows us to predict future college GPA for a set of
high school students with given characteristics.
3-7 Several Scenarios for Applying Multiple Regression
Now that we have covered the algebraic and statistical properties of OLS, it is a good time to catalog
different scenarios where unbiasedness of OLS can be established. In particular, we are interested in
situations verifying Assumptions MLR.1 and MLR.4, as these are the important population assumptions. Assumption MLR.2 concerns the sampling scheme, and the mild restrictions in Assumption
MLR.3 are rarely a concern.
The linearity assumption in MLR.1, where the error term u is additive, is always subject to criticism, although we know that it is not nearly as restrictive as it might seem because we can use transformations of both the explained and explanatory variables. Plus, the linear model is always a good
starting point, and often provides a suitable approximation. In any case, for the purposes of the following discussion the functional form issue is not critical.
58860_ch03_hr_066-116.indd 97
10/18/18 4:22 PM
98
PART 1 Regression Analysis with Cross-Sectional Data
3-7a Prediction
As suggested at the end of Section 3-6, sometimes we are interested in a pure prediction exercise,
where we hope to predict the outcome of a variable, y, given a set of observed variables, x1, x2, . . . , xk.
To continue the example previously mentioned, a college admissions officer might want to predict the
success of applicants—as measured by, say, future college GPA, y—based on information available
at the time of application. These variables, which include performance variables from high school
(GPA, kinds of classes taken, standardized test scores) and possibly family background, comprise
the explanatory variables. As described in Math Refresher B, the best predictor of y, as measured by
mean squared error, is the conditional expectation, E(y 0 x1, . . . , xk). If we assume a linear function for
the conditional expectation then
which is the same as writing
E(y 0 x1, . . . , xk) 5 b0 5 b1x1 1 ? ? ? 1 bkxk ,
y 5 b0 1 b1x1 1 ? ? ? 1 bkxk 1 u
E(u|x1, . . . , xk) 5 0.
In other words, MLR.4 is true by construction once we assume linearity. If we have a random sample
on the xj and y and we can rule out perfect collinearity, we can obtain unbiased estimators of the bj by
OLS. In the example of predicting future GPAs, we would obtain a sample of students who attended
the university so that we can observe their college GPAs. (Whether this provides a random sample
from the relevant population is an interesting question, but too advanced to discuss here. We do so in
Chapters 9 and 17.)
By estimating the bj we can also see which factors are most important for predicting future college success—as a way to fine tune our prediction model. But we do not yet have a formal way of
choosing which explanatory variables to include; that comes in the next chapter.
3-7b Efficient Markets
Efficient markets theories in economics, as well as some other theories, often imply that a single
variable acts as a “sufficient statistic” for predicting the outcome variable, y. For emphasis, call this
special predictor w. Then, given other observed factors, say x1, . . . , xk, we might want to test the
assumption
E(y 0 w, x) 5 E(y 0 w),
[3.63]
E(y 0 w, x) 5 b0 1 b1w 1 g1x1 1 ? ? ? 1 gkxk,
[3.64]
g1 5 g2 5 ? ? ? 5 gk 5 0.
[3.65]
where x is a shorthand for (x1, x2, . . . , xk). We can test (3.63) using a linear model for E(y 0 w, x):
where the slight change in notation is used to reflect the special status of w. In Chapter 4 we will learn
how to test whether all of the gj are zero:
Many efficient markets theories imply more than just (3.63). In addition, typically
E(y 0 w) 5 w,
which means that, in the linear equation (3.64), b0 5 0 and b1 5 1. Again, we will learn how to test
such restrictions in Chapter 4.
As a specific example, consider the sports betting market—say, for college football. The gambling markets produce a point spread, w 5 spread, which is determined prior to a game being played.
The spread typically varies a bit during the days preceding the game, but it eventually settles on some
58860_ch03_hr_066-116.indd 98
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
99
value. (Typically, the spread is in increments of 0.5.) The actually score differential in the game is
y 5 scorediff. Efficiency of the gambling market implies that
E(scoredifj 0 spread, x1, . . . , xk) 5 E(scorediff 0 spread) 5 spread,
where x1, . . . , xk includes any variables observable to the public prior to the game being played.
Examples include previous winning percentage, where the game is played, known injuries to key
players, and so on. The idea is that, because lots of money is involved in betting, the spread will move
until it incorporates all relevant information. Multiple regression can be used to test the efficient markets hypothesis because MLR.4 holds by construction once we assume a linear model:
y 5 b0 1 b1w 1 g1x1 1 ? ? ? 1 gkxk 1 u
E(u 0 w, x1, . . . , xk) 5 0,
[3.66]
[3.67]
where the explanatory variables are w, x1, . . . , xk.
Incidentally, it may be possible to think of a variable to be included among the xj that the market
has not incorporated into the spread. To be useful, it must be a variable that one can observe prior to
the game being played. Most tests of the efficiency of the gambling markets show that, except for
short aberrations, the market is remakably efficient.
3-7c Measuring the Tradeoff between Two Variables
Sometimes regression models are used not to predict, or to determe causality, but to simply measure
how an economic agent trades off one variable for another. Call these variables y and w. For example,
consider the population of K–12 teachers in a state in the United States. Let y be annual salary and
w be a measure of pension compensation. If teachers are indifferent between a dollar of salary and a
dollar of pension, then, on average, a one-dollar increase in pension compensation should be associated with a one-dollar fall in salary. In other words, only total compensation matters. Naturally, this is
a ceteris paribus question: all other relevant factors should be held fixed. In particular, we would expect
to see a positive correlation between salary and pension benefits because pension benefits are often
tied to salary. We want to know, for a given teacher, how does that teacher trade off one for the other.
Because we are simply measuring a tradeoff, it should not matter which variable we choose as
y and which we choose as w. However, functional form considerations can come into play. (We will
see this later in Example 4.10 in Chapter 4, where we study the salary-benefits tradeoff using aggregated data.) Once we have chosen y and w, and we have controls x 5 (x1, . . . , xk), we are, as in
Section 3-7b, interested in E(y 0 w, x). Assuming a linear model, we are exactly in the situation given in
equations (3.66) and (3.67). A key difference is that, assuming the xj properly control for differences in
individuals, the theory of a one-to-one tradeoff is b1 5 21, without restricting the intercept, b0. That
is quite different from the efficient markets hypothesis. Further, we include the xj to control for differences; we do not expect the yj to be zero, and we would generally have no interest in testing (3.65).
If we are not able to include sufficient controls in x then the estimated tradeoff coefficient, b, will
be biased (although the direction depends on what we think we have omitted). This is tantamount to
an omitted variable problem. For example, we may not have a suitable measure of teachers’ taste for
saving or amount of risk aversion.
3-7d Testing for Ceteris Paribus Group Differences
Another common application of multiple regression analysis is to test for differences among groups—
often, groups of people—once we account for other factors. In Section 2-7 we discussed the example
of estimating differences in hourly wage, wage, based on race, which is divided into white and other.
To this end, define a binary variable white. In Section 2-7a we noted that finding a difference in average wages across whites and nonwhites did not necessarily indicate wage discrimination because
other factors could contribute to such a differece.
58860_ch03_hr_066-116.indd 99
10/18/18 4:22 PM
100
PART 1 Regression Analysis with Cross-Sectional Data
Let x1, x2, . . . , xk denote other observable factors that can affect hourly wage—such as education,
workforce experience, and so on. Then we are interested in
E(wage 0 white, x1, . . . , xk).
If we have accounted for all factors in wage that should affect productivity, then wage differences by
race might be attributable to discrimination. In the simplest case, we would use a linear model:
E(wage 0 white, x1, . . . , xk) 5 b0 1 b1 white 1 g1x1 1 ? ? ? 1 gk xk,
[3.68]
where we are primarily interested in the coefficient b, which measures the difference in whites and
nonwhites given the same levels of the control variables, x1, x2, . . . , xk (education, experience, and
so on). For a general y and w, we again have (3.66) and (3.67) in force, and so MLR.4 holds by construction. OLS can be used to obtain an unbiased estimator of b (and the other coefficients). Problems
arise when we cannot include all suitable variables among the xj, in which case, again, we have an
omitted variable problem. In the case of testing for racial or gender discrimination, failure to control
for all relevant factors can cause systematic bias in estimating discrepancies due to discrimination.
3-7e Potential Outcomes, Treatment Effects, and Policy Analysis
For most practicing economists, the most exciting applications of multiple regression are in trying
to estimate causal effects of policy interventions. Do job training programs increase labor earnings?
By how much? Do school choice programs improve student outcomes? Does legalizing marijuana
increase crime rates?
We introduced the potential outcomes approach to studying policy questions in Section 2-7a. In
particular, we studied simple regression in the context of a binary policy intervention, using the notion
of counterfactual outcomes. In this section we change notation slightly, using w to denote the binary
intervention or policy indicator. As in Section 2-7a, for each unit in the population we imagine the
existence of the potential outcomes, y(0) and y(1)—representing different states of the world. If we
assume a constant treatment effect, say t, then we can write, for any unit i,
yi(1) 5 t 1 yi(0).
When the treatment effect can vary by i, the average treatment effect is
tate 5 E[yi(1) 2 yi(0)],
[3.69]
where the expectation is taken over the entire population.
For a random draw i, the outcome we observe, yi, can be written
yi 5 (1 2 wi)yi(0) 1 wi yi (1).
[3.70]
One of the important conclusions from Section 2-7a is that the simple regression of y on w (with
an intercept, as usual) is an unbiased estimator of tate only if we have random assignment of w—that is,
w is independent of [y(0), y(1)].
Random assignment is still pretty rare in business, economics, and other social sciences because true
experiments are still somewhat rare. Fortunately, if we can control variables—variables that help predict the potential outcomes and determine assignment into the treatment and control groups—we
can use multiple regression. Letting x again denote a set of control variables, consider the following
assumption:
w is independent of [y(0), y(1)] conditional on x.
[3.71]
For fairly obvious reasons, this assumption is called conditional independence, where it is important to note the variables in x that are in the conditioning set. In the treatment effects literature,
58860_ch03_hr_066-116.indd 100
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
101
(3.71) is also called unconfounded assignment or unconfoundedness conditional on x. The terms
ignorable assignment and ignorability are also used.
Assumption (3.71) has a simple interpretation. Think of partitioning the population based
on the observed variables in x. For concreteness, consider the job training program introduced in
Section 2-7a. There, w indicates whether a worker participates in a job training program, and y is
an outcome such as labor income. The elements in x include education, age, and past labor market
­history, such as earnings from the previous couple of years. Suppose that workers are more likely to
participate in the program the lower their education, age, and the worse their previous labor market
outcomes. Then, because education, age, and prior labor market history are very likely to predict y(0)
and y(1), random assignment does not hold. Nevertheless, once we group people by education, age,
and prior work history, it is possible that assignment is random. As a concrete example, consider the
group of people with 12 years of schooling who are 35 years old and who had average earnings of
$25,000 the past two years. What (3.71) requires is that within this group, assignment to the treatment
and control groups is random.
The more variables we observe prior to implementation of the program the more likely (3.71)
is to hold. If we observe no information to include in x then we are back to assuming pure random
assignment. Of course, it is always possible that we have not included the correct variables in x.
For example, perhaps everyone in a sample from the eligible population was administered a test to
measure intelligence, and assignment to the program is partly based on the score from the test. If we
observe the test score, we include it in x. If we cannot observe the test score, it must be excluded from
x and (3.71) would generally fail—although it could be “close” to being true if we have other good
controls in x.
How can we use (3.71) in multiple regression? Here we only consider the case of a constant treatment effect, t. Section 7-6 in Chapter 7 considers the more general case. Then, in the population,
y 5 y(0) 1 tw
and
E(y 0 w, x) 5 E[y(0) 0 w, x] 1 tw 5 E[y(0) 0 x] 1 tw,
[3.72]
where the second equality follows from conditional independence. Now assume that E[y(0) 0 x]
is linear,
Plugging in gives
E[y(0) 0 x] 5 a 1 xg.
E(y 0 w, x) 5 a 1 tw 1 xg 5 a 1 tw 1 g1x1 1 ? ? ? 1 xkxk.
[3.73]
As in several previous examples in this section, we are interested primarily in the coefficient on w,
which we have called t. The gj are of interest for logical consistency checks—for example, we should
expect more education to lead to higher earnings, on average—but the main role of the xj is to control
for differences across units.
In Chapter 7 we will cover treatment effects in more generality, including how to use multiple
regression when treatment effects vary by unit (individual in the job training case).
EXAMPLE 3.7
Evaluating a Job Training Program
The data in JTRAIN98 are on male workers that can be used to evaluate a job training program,
where the variable we would like to explain, y 5 earn98 is labor market earnings in 1998, the year
following the job training program (which took place in 1997). The earnings variable is measured
in thousands of dollars. The variable w 5 train is the binary participation (or “treatment”) indicator.
The participation in the job training program was partly based on past labor market outcomes and
58860_ch03_hr_066-116.indd 101
10/18/18 4:22 PM
102
PART 1 Regression Analysis with Cross-Sectional Data
is partly voluntary. Therefore, random assignment is unlikely to be a good assumption. As control
variables we use earnings in 1996 (the year prior to the program), years of schooling (educ), age, and
marital status (married). Like the training indicator, marital status is coded as a binary variable, where
married 5 1 means the man is married.
The simple regression estimates are
earn98 5 10.61 2 2.05 train
[3.74]
2
n 5 1,130, R 5 0.016
Because earns98 is measured in thousands of dollars, the coefficient on train, 22.05, shows that, on
average, those participating in the program earned $2,050 less than those who did not. The average
earnings for those who did not participate is gotten from the intercept, so $10,610.
Without random assignment, it is possible, even likely, that the negative (and large in magnitude) coefficient on train is a product of nonrandom selection into participation. This could be either
because men with poor earnings histories were more likely to be chosen or that such men are more
likely to participate if made eligible. We will not examine these propositions in detail here. Instead,
we add the four controls and perform a multiple regression:
earn98 5 4.67 1 2.41 train 1 . 373 earn96 1 . 363 educ 2 .181 age 1 2.48 married
n 5 1,130, R2 5 0.405
[3.75]
The change in the coefficient on train is remarkable: the program is now estimated to increase
earnings, on average, by $2,410. In other words, controlling for differences in preprogram earnings, education levels, age, and marital status produces a much different estimate than the simple
­difference-in-means estimate.
The signs of the coefficients on the control variables are not surprising. We expect earnings to be
positively correlated over time—so earns96 has a positive coefficient. Workers with more education
also earn more: about $363 for each additional year. The marriage effect is roughly as large as the job
training effect: ceteris paribus, married men earn, on average, about $2,480 more than their single
counterparts.
The predictability of the control variables is indicated by the R-squared in the multiple regression, R2 5 0.405. There is still much unexplained variation, but collectively the variables do a pretty
good job.
G o i n g F u rt h e r 3 . 5
Does it make sense to compare the intercepts in equations (3.74) and (3.75)?
Explain.
Before we end this section, a final remark is in
order. As with the other examples in this chapter, we
have not determined the statistical significance of the
estimates. We remedy this omission in Chapter 4,
where we learn how to test whether there is an effect
in the entire population, and also obtain confidence
intervals for the parameters, such as the average
treatment effect of a job training program.
Summary
1. The multiple regression model allows us to effectively hold other factors fixed while examining the
effects of a particular independent variable on the dependent variable. It explicitly allows the independent variables to be correlated.
2. Although the model is linear in its parameters, it can be used to model nonlinear relationships by
appropriately choosing the dependent and independent variables.
58860_ch03_hr_066-116.indd 102
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
103
3. The method of ordinary least squares is easily applied to estimate the multiple regression model. Each
slope estimate measures the partial effect of the corresponding independent variable on the dependent
variable, holding all other independent variables fixed.
4. R2 is the proportion of the sample variation in the dependent variable explained by the independent
variables, and it serves as a goodness-of-fit measure. It is important not to put too much weight on the
value of R2 when evaluating econometric models.
5. Under the first four Gauss-Markov assumptions (MLR.1 through MLR.4), the OLS estimators are
unbiased. This implies that including an irrelevant variable in a model has no effect on the unbiasedness of the intercept and other slope estimators. On the other hand, omitting a relevant variable causes
OLS to be biased. In many circumstances, the direction of the bias can be determined.
6. Under the five Gauss-Markov assumptions, the variance of an OLS slope estimator is given by
Var 1 b^ j 2 5 s2/ 3 SSTj 1 1 2 R2j 2 4 . As the error variance s2 increases, so does Var 1 b^ j 2 , while Var 1 b^ j 2
decreases as the sample variation in xj, SSTj, increases. The term R2j measures the amount of collinearity between xj and the other explanatory variables. As R2j approaches one, Var 1 b^ j 2 is unbounded.
7. Adding an irrelevant variable to an equation generally increases the variances of the remaining OLS
estimators because of multicollinearity.
8. Under the Gauss-Markov assumptions (MLR.1 through MLR.5), the OLS estimators are the best
­linear unbiased estimators (BLUEs).
9. Section 3-7 discusses the various ways that multiple regression analysis is used in economics and
other social sciences, including for prediction, testing efficient markets, estimating tradeoffs between
variables, and evaluating policy interventions. We will see examples of all such applications in the
remainder of the text.
10. Beginning in Chapter 4, we will use the standard errors of the OLS coefficients to compute confidence
intervals for the population parameters and to obtain test statistics for testing hypotheses about the
population parameters. Therefore, in reporting regression results we now include the standard errors
along with the associated OLS estimates. In equation form, standard errors are usually put in parentheses below the OLS estimates, and the same convention is often used in tables of OLS output.
The Gauss-Markov Assumptions
The following is a summary of the five Gauss-Markov assumptions that we used in this chapter. Remember,
the first four were used to establish unbiasedness of OLS, whereas the fifth was added to derive the usual
variance formulas and to conclude that OLS is best linear unbiased.
Assumption MLR.1 (Linear in Parameters)
The model in the population can be written as
y 5 b0 1 b1x1 1 b2x2 1 p 1 bkxk 1 u,
where b0, b1, . . . , bk are the unknown parameters (constants) of interest and u is an unobserved random
error or disturbance term.
Assumption MLR.2 (Random Sampling)
We have a random sample of n observations, 5 1 xi1, xi2, . . . , xik, yi 2 : i 5 1, 2, . . . , n 6 , following the population model in Assumption MLR.1.
Assumption MLR.3 (No Perfect Collinearity)
In the sample (and therefore in the population), none of the independent variables is constant, and there are
no exact linear relationships among the independent variables.
Assumption MLR.4 (Zero Conditional Mean)
The error u has an expected value of zero given any values of the independent variables. In other words,
E 1 u 0 x1, x2, . . . , xk 2 5 0.
58860_ch03_hr_066-116.indd 103
10/18/18 4:22 PM
104
PART 1 Regression Analysis with Cross-Sectional Data
Assumption MLR.5 (Homoskedasticity)
The error u has the same variance given any value of the explanatory variables. In other words,
Var 1 u 0 x1, . . . , xk 2 5 s2.
Key Terms
Best Linear Unbiased Estimator
(BLUE)
Biased Toward Zero
Ceteris Paribus
Conditional Independence
Degrees of Freedom (df)
Disturbance
Downward Bias
Endogenous Explanatory Variable
Error Term
Excluding a Relevant Variable
Exogenous Explanatory Variables
Explained Sum of Squares (SSE)
First Order Conditions
Frisch-Waugh Theorem
Gauss-Markov Assumptions
Gauss-Markov Theorem
Ignorable Assignment
Inclusion of an Irrelevant Variable
Intercept
Micronumerosity
Misspecification Analysis
Multicollinearity
Multiple Linear Regression (MLR)
Model
Multiple Regression Analysis
OLS Intercept Estimate
OLS Regression Line
OLS Slope Estimates
Omitted Variable Bias
Ordinary Least Squares
Overspecifying the Model
Partial Effect
Perfect Collinearity
Population Model
Residual
Residual Sum of Squares
Sample Regression
Function (SRF)
Slope Parameters
Standard Deviation of b^ j
Standard Error of b^ j
Standard Error of the
Regression (SER)
Sum of Squared Residuals (SSR)
Total Sum of Squares (SST)
True Model
Unconfounded Assignment
Underspecifying the Model
Upward Bias
Variance Inflation Factor (VIF)
Problems
1 Using the data in GPA2 on 4,137 college students, the following equation was estimated by OLS:
colgpa 5 1.392 2 .0135 hsperc 1 .00148 sat
n 5 4,137, R2 5 .273,
where colgpa is measured on a four-point scale, hsperc is the percentile in the high school graduating
class (defined so that, for example, hsperc = 5 means the top 5% of the class), and sat is the combined
math and verbal scores on the student achievement test.
(i) Why does it make sense for the coefficient on hsperc to be negative?
(ii) What is the predicted college GPA when hsperc = 20 and sat = 1,050?
(iii) Suppose that two high school graduates, A and B, graduated in the same percentile from high
school, but Student A’s SAT score was 140 points higher (about one standard deviation in the sample). What is the predicted difference in college GPA for these two students? Is the difference large?
(iv) Holding hsperc fixed, what difference in SAT scores leads to a predicted colgpa difference of
.50, or one-half of a grade point? Comment on your answer.
2 The data in WAGE2 on working men was used to estimate the following equation:
educ 5 10.36 2 .094 sibs 1 .131 meduc 1 .210 feduc
n 5 722, R2 5 .214,
where educ is years of schooling, sibs is number of siblings, meduc is mother’s years of schooling,
and feduc is father’s years of schooling.
(i) Does sibs have the expected effect? Explain. Holding meduc and feduc fixed, by how much does
sibs have to increase to reduce predicted years of education by one year? (A noninteger answer
is acceptable here.)
(ii) Discuss the interpretation of the coefficient on meduc.
58860_ch03_hr_066-116.indd 104
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
105
(iii) Suppose that Man A has no siblings, and his mother and father each have 12 years of education,
and Man B has no siblings, and his mother and father each have 16 years of education. What is
the predicted difference in years of education between B and A?
3 The following model is a simplified version of the multiple regression model used by Biddle and
Hamermesh (1990) to study the tradeoff between time spent sleeping and working and to look at
other factors affecting sleep:
sleep 5 b0 1 b1totwrk 1 b2educ 1 b3age 1 u,
where sleep and totwrk (total work) are measured in minutes per week and educ and age are measured
in years. (See also Computer Exercise C3 in Chapter 2.)
(i) If adults trade off sleep for work, what is the sign of b1?
(ii) What signs do you think b2 and b3 will have?
(iii) Using the data in SLEEP75, the estimated equation is
sleep 5 3,638.25 2 .148 totwrk 2 11.13 educ 1 2.20 age
n 5 706, R2 5 .113.
(iv)
(v)
If someone works five more hours per week, by how many minutes is sleep predicted to fall? Is
this a large tradeoff?
Discuss the sign and magnitude of the estimated coefficient on educ.
Would you say totwrk, educ, and age explain much of the variation in sleep? What other factors
might affect the time spent sleeping? Are these likely to be correlated with totwrk?
4 The median starting salary for new law school graduates is determined by
log 1 salary 2 5 b0 1 b1LSAT 1 b2GPA 1 b3log 1 libvol 2 1 b4log 1 cost 2
1 b5rank 1 u,
where LSAT is the median LSAT score for the graduating class, GPA is the median college GPA for the
class, libvol is the number of volumes in the law school library, cost is the annual cost of attending law
school, and rank is a law school ranking (with rank 5 1 being the best).
(i) Explain why we expect b5 # 0.
(ii) What signs do you expect for the other slope parameters? Justify your answers.
(iii) Using the data in LAWSCH85, the estimated equation is
log 1 salary 2 5 8.34 1 .0047 LSAT 1 .248 GPA 1 .095 log 1 libvol 2
1 .038 log 1 cost 2 2 .0033 rank
n 5 136, R2 5 .842.
(iv)
(v)
What is the predicted ceteris paribus difference in salary for schools with a median GPA different by one point? (Report your answer as a percentage.)
Interpret the coefficient on the variable log(libvol).
Would you say it is better to attend a higher ranked law school? How much is a difference in
ranking of 20 worth in terms of predicted starting salary?
5 In a study relating college grade point average to time spent in various activities, you distribute a survey to several students. The students are asked how many hours they spend each week in four activities: studying, sleeping, working, and leisure. Any activity is put into one of the four categories, so that
for each student, the sum of hours in the four activities must be 168.
(i) In the model
GPA 5 b0 1 b1study 1 b2sleep 1 b3work 1 b4leisure 1 u,
(ii)
58860_ch03_hr_066-116.indd 105
does it make sense to hold sleep, work, and leisure fixed, while changing study?
Explain why this model violates Assumption MLR.3.
10/18/18 4:22 PM
106
PART 1 Regression Analysis with Cross-Sectional Data
(iii) How could you reformulate the model so that its parameters have a useful interpretation and it
satisfies Assumption MLR.3?
6 Consider the multiple regression model containing three independent variables, under Assumptions
MLR.1 through MLR.4:
y 5 b0 1 b1x1 1 b2x2 1 b3x3 1 u.
You are interested in estimating the sum of the parameters on x1 and x2; call this u 1 5 b0 1 b1.
(i) Show that u^ 1 5 b^ 1 1 b^ 2 is an unbiased estimator of u 1.
^ 22.
(ii) Find Var 1 u^ 1 2 in terms of Var 1 b^ 1 2 ,Var 1 b^ 2 2 , and Corr 1 b^ 1, b
7 Which of the following can cause OLS estimators to be biased?
(i) Heteroskedasticity.
(ii) Omitting an important variable.
(iii) A sample correlation coefficient of .95 between two independent variables both included in the
model.
8 Suppose that average worker productivity at manufacturing firms (avgprod) depends on two factors,
average hours of training (avgtrain) and average worker ability (avgabil):
avgprod 5 b0 1 b1avgtrain 1 b2avgabil 1 u.
Assume that this equation satisfies the Gauss-Markov assumptions. If grants have been given to firms
whose workers have less than average ability, so that avgtrain and avgabil are negatively correlated,
| obtained from the simple regression of avgprod on avgtrain?
what is the likely bias in b
1
9 The following equation describes the median housing price in a community in terms of amount of pollution (nox for nitrous oxide) and the average number of rooms in houses in the community (rooms):
(i)
(ii)
log 1 price 2 5 b0 1 b1log 1 nox 2 1 b2rooms 1 u.
What are the probable signs of b1 and b2? What is the interpretation of b1? Explain.
Why might nox [or more precisely, log(nox)] and rooms be negatively correlated? If this is the
case, does the simple regression of log(price) on log(nox) produce an upward or a downward
biased estimator of b1?
(iii) Using the data in HPRICE2, the following equations were estimated:
log 1 price 2 5 11.71 2 1.043 log 1 nox 2 , n 5 506, R2 5 .264.
log 1 price 2 5 9.23 2 .718 log 1 nox 2 1 .306 rooms, n 5 506, R2 5 .514.
Is the relationship between the simple and multiple regression estimates of the elasticity of price
with respect to nox what you would have predicted, given your answer in part (ii)? Does this
mean that 2.718 is definitely closer to the true elasticity than 21.043?
10 Suppose that you are interested in estimating the ceteris paribus relationship between y and x1. For this
purpose, you can collect data on two control variables, x2 and x3. (For concreteness, you might think
of y as final exam score, x1 as class attendance, x2 as GPA up through the previous semester, and x3 as
| be the simple regression estimate from y on x and let b^ be the multiple
SAT or ACT score.) Let b
1
1
1
regression estimate from y on x1, x2, x3.
(i) If x1 is highly correlated with x2 and x3 in the sample, and x2 and x3 have large partial effects
| and b^ to be similar or very different? Explain.
on y, would you expect b
1
1
| and b^
(ii) If x1 is almost uncorrelated with x2 and x3, but x2 and x3 are highly correlated, will b
1
1
tend to be similar or very different? Explain.
(iii) If x1 is highly correlated with x2 and x3, and x2 and x3 have small partial effects on y, would you
| 2 or se 1 b^ 2 to be smaller? Explain.
expect se 1 b
1
1
(iv) If x1 is almost uncorrelated with x2 and x3, x2 and x3 have large partial effects on y, and x2 and x3
| 2 or se 1 b^ 2 to be smaller? Explain.
are highly correlated, would you expect se 1 b
1
1
58860_ch03_hr_066-116.indd 106
10/18/18 4:22 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
107
11 Suppose that the population model determining y is
y 5 b0 1 b1x1 1 b2x2 1 b3x3 1 u,
and this model satisfies Assumptions MLR.1 through MLR.4. However, we estimate the model that
|,b
| , and b
| be the OLS estimators from the regression of y on x and x . Show that the
omits x3. Let b
0
1
2
1
2
|
expected value of b1 (given the values of the independent variables in the sample) is
a r^ i1xi3
n
| 2 5b 1b
E1b
1
1
3
i51
n
a r^ i1
2
,
i51
| comes
where the r^ i1 are the OLS residuals from the regression of x1 on x2. [Hint: The formula for b
1
from equation (3.22). Plug yi 5 b0 1 b1xi1 1 b2xi2 1 b3xi3 1 ui into this equation. After some
algebra, take the expectation treating xi3 and r^ i1 as nonrandom.]
12 The following equation represents the effects of tax revenue mix on subsequent employment growth
for the population of counties in the United States:
growth 5 b0 1 b1shareP 1 b2shareI 1 b3shareS 1 other factors,
where growth is the percentage change in employment from 1980 to 1990, shareP is the share of property taxes in total tax revenue, shareI is the share of income tax revenues, and shareS is the share of
sales tax revenues. All of these variables are measured in 1980. The omitted share, shareF, includes
fees and miscellaneous taxes. By definition, the four shares add up to one. Other factors would include
expenditures on education, infrastructure, and so on (all measured in 1980).
(i) Why must we omit one of the tax share variables from the equation?
(ii) Give a careful interpretation of b1.
13 (i)Consider the simple regression model y 5 b0 1 b1x 1 u under the first four Gauss-Markov
assumptions. For some function g(x), for example g 1 x 2 5 x2 or g 1 x 2 5 log 1 1 1 x2 2 , define
zi 5 g 1 xi 2 . Define a slope estimator as
n
| 5 n 1z 2 z2y
b
aa i
1
i b ^ a a 1 zi 2 z 2 xi b.
i51
(ii)
i51
| is linear and unbiased. Remember, because E(u|x) = 0, you can treat both x and z
Show that b
1
i
i
as nonrandom in your derivation.
Add the homoskedasticity assumption, MLR.5. Show that
n
| 2 5 s2 n 1 z 2 z 2 2
Var 1 b
aa i
b ^ a a 1 z i 2 z 2 xi b .
1
2
i51
i51
| 2 , where b^ is the
(iii) Show directly that, under the Gauss-Markov assumptions, Var 1 b^ 1 2 # Var 1 b
1
1
OLS estimator. [Hint: The Cauchy-Schwartz inequality in Math Refresher B implies that
an 21 a 1 zi 2 z 2 1 xi 2 x 2 b # an 21 a 1 zi 2 z 2 2 b an 21 a 1 xi 2 x 2 2 b;
n
2
i51
n
n
i51
i51
notice that we can drop x from the sample covariance.]
14 Suppose you have a sample of size n on three variables, y, x1, and x2, and you are primarily interested
| be the coefficient on x from the simple regression and b^ the coefficient
in the effect of x1 on y. Let b
1
1
1
on x1 from the regression y on x1, x2. The standard errors reported by any regression package are
| 2 5
se 1 b
1
58860_ch03_hr_066-116.indd 107
|
s
"SST1
10/18/18 4:22 PM
108
PART 1 Regression Analysis with Cross-Sectional Data
se 1 b^ 1 2 5
s^
"SST1
? "VIF1,
| is the SER from the simple regression, s^ is the SER from the multiple regression,
where s
VIF1 5 1/ 1 1 2 R21 2 , and R21 is the R-squared from the regression of x1 on x2. Explain why
| 2.
se 1 b^ 12 can be smaller or larger than se 1 b
1
15 The following estimated equations use the data in MLB1, which contains information on major league
baseball salaries. The dependent variable, lsalary, is the log of salary. The two explanatory variables
are years in the major leagues (years) and runs batted in per year (rbisyr):
lsalary 5 12.373 1 .1770 years
1 .098 2 1 .0132 2
n 5 353, SSR 5 326.196, SER 5 .964, R2 5 .337
lsalary 5 11.861 1 .0904 years 1 .0302 rbisyr
1 .084 2 1 .0118 2
1 .0020 2
n 5 353, SSR 5 198.475, SER 5 .753, R2 5 .597
(i)
How many degrees of freedom are in each regression? Why is the SER smaller in the second
regression than the first?
(ii) The sample correlation coefficient between years and rbisyr is about 0.487. Does this make
sense? What is the variance inflation factor (there is only one) for the slope coefficients in the
multiple regression? Would you say there is little, moderate, or strong collinearity between years
and rbisyr?
(iii) How come the standard error for the coefficient on years in the multiple regression is lower than
its counterpart in the simple regression?
16 The following equations were estimated using the data in LAWSCH85:
lsalary 5 9.90 2 .0041 rank 1 .294 GPA
1 .24 2 1 .0003 2
1 .069 2
n 5 142, R2 5 .8238
lsalary 5 9.86 2 .0038 rank 1 .295 GPA 1 .00017 age
1 .29 2 1 .0004 2
1 .083 2
1 .00036 2
n 5 99, R2 5 .8036
How can it be that the R-squared is smaller when the variable age is added to the equation?
17 Consider an estimated equation for workers earning an hourly wage, wage,where educ, years of
schooling, and exper, actual years in the workforce, are measured in years. The dependent variable is
lwage 5 log(wage):
lwage 5 0.532 1 .094 educ 1 .026 exper
n 5 932, R2 5 0.188
Suppose that getting one more year of education necessarily reduces workforce experience by one
year. What is the estimated percentage change in wage from getting one more year of schooling?
18 The potential outcomes framework in Section 3-7e can be extended to more than two potential outcomes. In fact, we can think of the policy variable, w, as taking on many different values, and then y(w)
denotes the outcome for policy level w. For concreteness, suppose w is the dollar amount of a grant that
can be used for purchasing books and electronics in college, y(w) is a measure of college performance,
58860_ch03_hr_066-116.indd 108
10/18/18 4:23 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
109
such as grade point average. For example, y(0) is the resulting GPA if the student receives no grant
and y(500) is the resulting GPA if the grant amount is $500.
For a random draw i, we observe the grant level, wi $ 0 and yi 5 y(wi). As in the binary program
evaluation case, we observe the policy level, wi , and then only the outcome associated with that level.
(i) Suppose a linear relationship is assumed:
y(w) 5 a 1 bw 1 v(0)
where y(0) 5 a 1 v. Further, assume that for all i, wi is independent of vi. Show that for each i
we can write
yi 5 a 1 bwi 1 vi
E(vi 0 wi) 5 0.
(ii)
In the setting of part (i), how would you estimate b (and a) given a random sample? Justify
your answer.
(iii) Now suppose that wi is possibly correlated with vi, but for a set of observed variables xij,
E(vi 0 wi, xi1, . . . , xik) 5 E(vi 0 xi1, . . . , xik) 5 h 1 g1xi1 1 ? ? ? 1 gkxik.
The first equality holds if wi is independent of vi conditional on (xi1, . . . , xik ) and the second
equality assumes a linear relationship. Show that we can write
(iv)
yi 5 c 1 bwi 1 g1xi1 1 ? ? ? 1 gkxik 1 ui
E(ui 0 wi, xi1, . . . , xik) 5 0.
What is the intercept c?
How would you estimate b (along with c and the gj) in part (iii)? Explain.
Computer Exercises
C1 A problem of interest to health officials (and others) is to determine the effects of smoking during
pregnancy on infant health. One measure of infant health is birth weight; a birth weight that is too
low can put an infant at risk for contracting various illnesses. Since factors other than cigarette smoking that affect birth weight are likely to be correlated with smoking, we should take those factors into
account. For example, higher income generally results in access to better prenatal care, as well as better nutrition for the mother. An equation that recognizes this is
bwght 5 b0 1 b1cigs 1 b2faminc 1 u.
(i)
(ii)
What is the most likely sign for b2?
Do you think cigs and faminc are likely to be correlated? Explain why the correlation might be
positive or negative.
(iii) Now, estimate the equation with and without faminc, using the data in BWGHT. Report the
results in equation form, including the sample size and R-squared. Discuss your results, focusing on whether adding faminc substantially changes the estimated effect of cigs on bwght.
C2 Use the data in HPRICE1 to estimate the model
price 5 b0 1 b1sqrft 1 b2bdrms 1 u,
where price is the house price measured in thousands of dollars.
(i) Write out the results in equation form.
(ii) What is the estimated increase in price for a house with one more bedroom, holding square
footage constant?
(iii) What is the estimated increase in price for a house with an additional bedroom that is
140 square feet in size? Compare this to your answer in part (ii).
58860_ch03_hr_066-116.indd 109
10/18/18 4:23 PM
110
PART 1 Regression Analysis with Cross-Sectional Data
(iv)
What percentage of the variation in price is explained by square footage and number
of bedrooms?
(v) The first house in the sample has sqrft 5 2,438 and bdrms 5 4. Find the predicted selling price
for this house from the OLS regression line.
(vi) The actual selling price of the first house in the sample was $300,000 (so price 5 300). Find
the residual for this house. Does it suggest that the buyer underpaid or overpaid for the house?
C3 The file CEOSAL2 contains data on 177 chief executive officers and can be used to examine
the effects of firm performance on CEO salary.
(i) Estimate a model relating annual salary to firm sales and market value. Make the model
of the constant elasticity variety for both independent variables. Write the results out
in equation form.
(ii) Add profits to the model from part (i). Why can this variable not be included in logarithmic
form? Would you say that these firm performance variables explain most of the variation
in CEO salaries?
(iii) Add the variable ceoten to the model in part (ii). What is the estimated percentage return for
another year of CEO tenure, holding other factors fixed?
(iv) Find the sample correlation coefficient between the variables log(mktval) and profits. Are these
variables highly correlated? What does this say about the OLS estimators?
C4 Use the data in ATTEND for this exercise.
(i) Obtain the minimum, maximum, and average values for the variables atndrte, priGPA, and ACT.
(ii) Estimate the model
atndrte 5 b0 1 b1priGPA 1 b2ACT 1 u,
and write the results in equation form. Interpret the intercept. Does it have a useful meaning?
(iii) Discuss the estimated slope coefficients. Are there any surprises?
(iv) What is the predicted atndrte if priGPA 5 3.65 and ACT 5 20? What do you make of this
result? Are there any students in the sample with these values of the explanatory variables?
(v) If Student A has priGPA 5 3.1 and ACT 5 21 and Student B has priGPA 5 2.1 and
ACT 5 26, what is the predicted difference in their attendance rates?
C5 Confirm the partialling out interpretation of the OLS estimates by explicitly doing the partialling out
for Example 3.2. This first requires regressing educ on exper and tenure and saving the residuals, r^ 1.
Then, regress log(wage) on r^ 1. Compare the coefficient on r^ 1 with the coefficient on educ in the regression of log(wage) on educ, exper, and tenure.
C6 Use the data set in WAGE2 for this problem. As usual, be sure all of the following regressions contain
an intercept.
(i) Run a simple regression of IQ on educ to obtain the slope coefficient, say, |
d 1.
|.
(ii) Run the simple regression of log(wage) on educ, and obtain the slope coefficient, b
1
(iii) Run the multiple regression of log(wage) on educ and IQ, and obtain the slope coefficients,
b^ 1 and b^ 2, respectively.
| 5 b^ 1 b^ |
(iv) Verify that b
1
1
2 d 1.
C7 Use the data in MEAP93 to answer this question.
(i) Estimate the model
math10 5 b0 1 b1log 1 expend 2 1 b2lnchprg 1 u,
(ii)
58860_ch03_hr_066-116.indd 110
and report the results in the usual form, including the sample size and R-squared. Are the signs
of the slope coefficients what you expected? Explain.
What do you make of the intercept you estimated in part (i)? In particular, does it make sense to
set the two explanatory variables to zero? [Hint: Recall that log 1 1 2 5 0.]
10/18/18 4:23 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
111
(iii) Now run the simple regression of math10 on log(expend), and compare the slope coefficient
with the estimate obtained in part (i). Is the estimated spending effect now larger or smaller
than in part (i)?
(iv) Find the correlation between lexpend 5 log 1 expend 2 and lnchprg. Does its sign make sense to you?
(v) Use part (iv) to explain your findings in part (iii).
C8 Use the data in DISCRIM to answer this question. These are zip code–level data on prices for various items at fast-food restaurants, along with characteristics of the zip code population, in New Jersey
and Pennsylvania. The idea is to see whether fast-food restaurants charge higher prices in areas with a
larger concentration of blacks.
(i) Find the average values of prpblck and income in the sample, along with their standard deviations. What are the units of measurement of prpblck and income?
(ii) Consider a model to explain the price of soda, psoda, in terms of the proportion of the population that is black and median income:
psoda 5 b0 1 b1prpblck 1 b2income 1 u.
Estimate this model by OLS and report the results in equation form, including the sample size
and R-squared. (Do not use scientific notation when reporting the estimates.) Interpret the coefficient on prpblck. Do you think it is economically large?
(iii) Compare the estimate from part (ii) with the simple regression estimate from psoda on prpblck.
Is the discrimination effect larger or smaller when you control for income?
(iv) A model with a constant price elasticity with respect to income may be more appropriate.
Report estimates of the model
log 1 psoda 2 5 b0 1 b1prpblck 1 b2log 1 income 2 1 u.
If prpblck increases by .20 (20 percentage points), what is the estimated percentage change
in psoda? (Hint: The answer is 2.xx, where you fill in the “xx.”)
(v) Now add the variable prppov to the regression in part (iv). What happens to b^ prpblck?
(vi) Find the correlation between log(income) and prppov. Is it roughly what you expected?
(vii) Evaluate the following statement: “Because log(income) and prppov are so highly correlated,
they have no business being in the same regression.”
C9 Use the data in CHARITY to answer the following questions:
(i) Estimate the equation
gift 5 b0 1 b1mailsyear 1 b2giftlast 1 b3propresp 1 u
by OLS and report the results in the usual way, including the sample size and R-squared.
How does the R-squared compare with that from the simple regression that omits giftlast
and propresp?
(ii) Interpret the coefficient on mailsyear. Is it bigger or smaller than the corresponding simple
regression coefficient?
(iii) Interpret the coefficient on propresp. Be careful to notice the units of measurement of propresp.
(iv) Now add the variable avggift to the equation. What happens to the estimated effect
of mailsyear?
(v) In the equation from part (iv), what has happened to the coefficient on giftlast? What do you
think is happening?
C10 Use the data in HTV to answer this question. The data set includes information on wages, education,
parents’ education, and several other variables for 1,230 working men in 1991.
(i) What is the range of the educ variable in the sample? What percentage of men completed
twelfth grade but no higher grade? Do the men or their parents have, on average, higher levels
of education?
58860_ch03_hr_066-116.indd 111
10/18/18 4:23 PM
112
PART 1 Regression Analysis with Cross-Sectional Data
(ii)
Estimate the regression model
educ 5 b0 1 b1motheduc 1 b2fatheduc 1 u
by OLS and report the results in the usual form. How much sample variation in educ is
­explained by parents’ education? Interpret the coefficient on motheduc.
(iii) Add the variable abil (a measure of cognitive ability) to the regression from part (ii), and report
the results in equation form. Does “ability” help to explain variations in education, even after
controlling for parents’ education? Explain.
(iv) (Requires calculus) Now estimate an equation where abil appears in quadratic form:
educ 5 b0 1 b1motheduc 1 b2fatheduc 1 b3abil 1 b4abil2 1 u.
Using the estimates b^ 3 and b^ 4, use calculus to find the value of abil, call it abil*, where educ
is minimized. (The other coefficients and values of parents’ education variables have no effect;
we are holding parents’ education fixed.) Notice that abil is measured so that negative values
are permissible. You might also verify that the second derivative is positive so that you do
indeed have a minimum.
(v) Argue that only a small fraction of men in the sample have “ability” less than the value calculated in part (iv). Why is this important?
(vi) If you have access to a statistical program that includes graphing capabilities, use the estimates
in part (iv) to graph the relationship between the predicted education and abil. Set motheduc and
fatheduc at their average values in the sample, 12.18 and 12.45, respectively.
C11 Use the data in MEAPSINGLE to study the effects of single-parent households on student math performance. These data are for a subset of schools in southeast Michigan for the year 2000. The socioeconomic variables are obtained at the ZIP code level (where ZIP code is assigned to schools based on
their mailing addresses).
(i) Run the simple regression of math4 on pctsgle and report the results in the usual format.
Interpret the slope coefficient. Does the effect of single parenthood seem large or small?
(ii) Add the variables lmedinc and free to the equation. What happens to the coefficient on pctsgle?
Explain what is happening.
(iii) Find the sample correlation between lmedinc and free. Does it have the sign you expect?
(iv) Does the substantial correlation between lmedinc and free mean that you should drop one from
the regression to better estimate the causal effect of single parenthood on student performance?
Explain.
(v) Find the variance inflation factors (VIFs) for each of the explanatory variables appearing
in the regression in part (ii). Which variable has the largest VIF? Does this knowledge
affect the model you would use to study the causal effect of single parenthood on math
performance?
C12 The data in ECONMATH contain grade point averages and standardized test scores, along with
performance in an introductory economics course, for students at a large public university. The variable to be explained is score, the final score in the course measured as a percentage.
(i) How many students received a perfect score for the course? What was the average score? Find
the means and standard deviations of actmth and acteng, and discuss how they compare.
(ii) Estimate a linear equation relating score to colgpa, actmth, and acteng, where colgpa is
measured at the beginning of the term. Report the results in the usual form.
(iii) Would you say the math or English ACT score is a better predictor of performance in the
economics course? Explain.
(iv) Discuss the size of the R-squared in the regression.
58860_ch03_hr_066-116.indd 112
10/18/18 4:23 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
113
C13 Use the data in GPA1 to answer this question. We can compare multiple regression estimates, where
we control for student achievement and background variables, and compare our findings with the
­difference-in-means estimate in Computer Exercise C11 in Chapter 2.
(i) In the simple regression equation
colGPA 5 b0 1 b1PC 1 u
obtain b^ 0 and b^ 1. Interpret these estimates.
(ii) Now add the controls hsGPA and ACT—that is, run the regression colGPA on PC, hsGPA, and
ACT. Does the coefficient on PC change much from part (ii)? Does b^ hsGPA make sense?
(iii) In the estimation from part (ii), what is worth more: Owning a PC or having 10 more points on
the ACT score?
(iv) Now to the regression in part (ii) add the two binary indicators for the parents being college
graduates. Does the estimate of b1 change much from part (ii)? How much variation are you
explaining in colGPA?
(v) Suppose someone looking at your regression from part (iv) says to you, “The variables hsGPA
and ACT are probably pretty highly correlated, so you should drop one of them from the
­regression.” How would you respond?
Appendix 3A
3A.1 Derivation of the First Order Conditions in Equation (3.13)
The analysis is very similar to the simple regression case. We must characterize the solutions to
the problem
min a 1 yi 2 b0 2 b1xi1 2 p 2 bkxik 2 2.
n
b0, b1p , bk i51
Taking the partial derivatives with respect to each of the bj (see Math Refresher A), evaluating them
at the solutions, and setting them equal to zero gives
22 a 1 yi 2 b^ 0 2 b^ 1xi1 2 p 2 b^ kxik 2 5 0
n
i51
22 a xij 1 yi 2 b^ 0 2 b^ 1xi1 2 p 2 b^ kxik 2 5 0, for all j 5 1, . . . , k.
n
i51
Canceling the 22 gives the first order conditions in (3.13).
3A.2 Derivation of Equation (3.22)
To derive (3.22), write xi1 in terms of its fitted value and its residual from the regression of x1 on
x2, . . . , xk: xi1 5 x^ i1 1 r^ i1, for all i 5 1, . . . , n. Now, plug this into the second equation in (3.13):
p 2 b^ kxik 2 5 0.
^
^
a 1 x^ i1 1 r^ i1 2 1 yi 2 b0 2 b1xi1 2
n
i51
[3.76]
By the definition of the OLS residual u^ i, because x^ i1 is just a linear function of the explanatory variables xi2, . . . , xik, it follows that g ni51 x^ i1u^ i 5 0. Therefore, equation (3.76) can be expressed as
p 2 b^ kxik 2 5 0.
^
^
a r^ i1 1 yi 2 b0 2 b1xi1 2
n
i51
58860_ch03_hr_066-116.indd 113
[3.77]
10/18/18 4:23 PM
114
PART 1 Regression Analysis with Cross-Sectional Data
Because the r^ i1 are the residuals from regressing x1 on x2, c, xk , g i51 xijr^ i1 5 0, for all
n
j 5 2, . . . , k. Therefore, (3.77) is equivalent to g i51 r^ i1 1 yi 2 b^ 1xi1 2 5 0. Finally, we use the fact that
n
g i51 x^ i1r^ i1 5 0, which means that b^ 1 solves
n
^
a r^ i1 1 yi 2 b1r^ i1 2 5 0.
n
Now, straightforward algebra gives (3.22), provided, of course, that g i51 r^ 2i1 . 0; this is ensured by
Assumption MLR.3.
i51
n
3A.3 Proof of Theorem 3.1
We prove Theorem 3.1 for b^ 1; the proof for the other slope parameters is virtually identical. (See
Advanced Treatment E for a more succinct proof using matrices.) Under Assumption MLR.3, the OLS
estimators exist, and we can write b^ 1 as in (3.22). Under Assumption MLR.1, we can write yi as in
n
n
(3.32); substitute this for yi in (3.22). Then, using g i51r^ i1 5 0, g i51 xijr^ i1 5 0, for all j 5 2, . . . , k,
n
n
2
and g i51 xi1r^ i1 5 g i51r^ i1, we have
b^ 1 5 b1 1 a a r^ i1ui b ^ a a r^ 2i1 b.
n
n
i51
i51
[3.78]
Now, under Assumptions MLR.2 and MLR.4, the expected value of each ui, given all independent
variables in the sample, is zero. Because the r^ i1 are just functions of the sample independent variables, it follows that
E 1 b^ 1 0 X 2 5 b1 1 a a r^ i1E 1 ui 0 X 2 b ^ a a r^ 2i1 b n
n
i51
i51
5 b1 1 a a r^ i1 # 0b ^ a a r^ 2i1 b 5 b1,
n
n
i51
i51
where X denotes the data on all independent variables and E 1 b^ 1 0 X 2 is the expected value of b^ 1, given
xi1, . . . , xik, for all i 5 1, . . . , n. This completes the proof.
3A.4 General Omitted Variable Bias
We can derive the omitted variable bias in the general model in equation (3.31) under the first four
Gauss-Markov assumptions. In particular, let the b^ j, j 5 0, 1, . . . , k be the OLS estimators from the
| , j 5 0, 1, . . . , k 2 1 be the OLS
regression using the full set of explanatory variables. Let the b
j
|
estimators from the regression that leaves out xk. Let d j, j 5 1, . . . , k 2 1 be the slope coefficient
on xj in the auxiliary regression of xik on xi1, xi2, . . . xi,k21, i 5 1, . . . , n. A useful fact is that
| 5 b^ 1 b^ |
b
d .[3.79]
j
j
k j
This shows explicitly that, when we do not control for xk in the regression, the estimated partial
effect of xj equals the partial effect when we include xk plus the partial effect of xk on y^ times the
partial relationship between the omitted variable, xk, and xj, j , k. Conditional on the entire set of
explanatory variables, X, we know that the b^ j are all unbiased for the corresponding bj, j 5 1, . . . , k.
Further, because |
d j is just a function of X, we have
| 0 X 2 5 E 1 b^ 0 X 2 1 E 1 b^ 0 X 2 |
E1b
d j
j
j
k
|
5b 1b d.
j
58860_ch03_hr_066-116.indd 114
[3.80]
k j
10/18/18 4:23 PM
CHAPTER 3 Multiple Regression Analysis: Estimation
115
| is biased for b unless b 5 0—in which case x has no partial effect
Equation (3.80) shows that b
j
j
k
k
|
in the population—or d j equals zero, which means that xik and xij are partially uncorrelated in the
sample. The key to obtaining equation (3.80) is equation (3.79). To show equation (3.79), we can use
| is the slope coefficient
equation (3.22) a couple of times. For simplicity, we look at j 5 1. Now, b
1
in the simple regression of yi on |
r i1, i 5 1, . . . , n, where the |r i1 are the OLS residuals from the
| : g n |r y .
regression of xi1 on xi2, xi3, . . . , xi,k21. Consider the numerator of the expression for b
i51 i1 i
1
But for each i, we can write yi 5 b^ 0 1 b^ 1xi1 1 p 1 b^ kxik 1 u^ i and plug in for yi . Now, by
properties of the OLS residuals, the |
r i1 have zero sample average and are uncorrelated with
xi2, xi3, . . . , xi,k21 in the sample. Similarly, the u^ i have zero sample average and zero sample correlation with xi1, xi2, c, xik. It follows that the |r i1 and u^ i are uncorrelated in the sample (because the |r i1
are just linear combinations of xi1, xi2, c, xi,k21). So
r i1yi 5 b^ 1 a a |r i1xi1 b 1 b^ k a a |r i1xik b.
a|
n
n
n
[3.81]
n
n
| . Therefore, we have shown that
Now, g i51 |r i1xi1 5 g i51 |
r 2i1, which is also the denominator of b
1
i51
i51
i51
n
n
| 5 b^ 1 b^
|
|2
b
1
1
k a a r i1xik b ^ a a r i1, b i51
i51
5 b^ 1 1 b^ k|
d 1.
This is the relationship we wanted to show.
3A.5 Proof of Theorem 3.2
Again, we prove this for j 5 1. Write b^ 1 as in equation (3.78). Now, under MLR.5, Var 1 ui 0 X 2 5 s2,
for all i 5 1, c, n. Under random sampling, the ui are independent, even conditional on X, and the
r^ i1 are nonrandom conditional on X. Therefore,
Var 1 b^ 1 0 X 2 5 a a r^ 2i1 Var 1 ui 0 X 2 b ^ a a r^ 2i1 b
n
n
i51
i51
2
5 a a r^ 2i1s2 b ^ a a r^ 2i1 b 5 s2^ a a r^ 2i1 b.
n
n
2
Now, because g ni51 r^ 2i1 is the sum of squared residuals
g ni51 r^ 2i1 5 SST1 1 1 2 R21 2 . This completes the proof.
i51
i51
n
i51
from regressing x1 on x2, . . . , xk ,
3A.6 Proof of Theorem 3.4
| of b , Var 1 b
| 2 $ Var 1 b^ 2 , where b^ is the
We show that, for any other linear unbiased estimator b
1
1
1
1
1
OLS estimator. The focus on j 5 1 is without loss of generality.
| as in equation (3.60), we can plug in for y to obtain
For b
1
i
| 5 b n w 1 b n w x 1 b n w x 1 p 1 b n w x 1 n w u.
b
a i1 i
1
0 a i1
1 a i1 i1
2 a i1 i2
k a i1 ik
i51
i51
i51
i51
i51
Now, because the wi1 are functions of the xij,
| 0X2 5 b n w 1 b n w x 1 b n w x 1 p 1 b n w x 1 n w E1u 0X2
E1b
a i1
1
0 a i1
1 a i1 i1
2 a i1 i2
k a i1 ik
i
i51
n
i51
n
i51
n
i51
i51
i51
i51
n
5 b0 a wi1 1 b1 a wi1xi1 1 b2 a wi1xi2 1 p 1 bk a wi1xik
58860_ch03_hr_066-116.indd 115
i51
i51
10/18/18 4:23 PM
116
PART 1 Regression Analysis with Cross-Sectional Data
| 0 X 2 to
because E 1 ui 0 X 2 5 0, for all i 5 1, . . . , n under MLR.2 and MLR.4. Therefore, for E 1 b
1
equal b1 for any values of the parameters, we must have
a wi1 5 0, a wi1xi1 5 1, a wi1xij 5 0, j 5 2, . . . , k.
n
n
n
i51
i51
i51
[3.82]
Now, let r^ i1 be the residuals from the regression of xi1 on xi2, . . . , xik. Then, from (3.82), it follows that
a wi1ri1 5 1
n
n
g i51wi1x^ i1
[3.83]
i51
| 0 X 2 and
5 0. Now, consider the difference between Var 1 b
because xi1 5 x^ i1 1 r^ i1 and
1
^
Var 1 b1 0 X 2 under MLR.1 through MLR.5:
s2 a w2i1 2 s2^ a a r^ 2i1 b.[3.84]
n
n
i51
i51
Because of (3.83), we can write the difference in (3.84), without s2, as
2
a wi1 2 a a wi1r^ i1 b
But (3.85) is simply
n
n
i51
i51
2
2
^ a i51
a r^ i1 b.
2
a 1 wi1 2 g^ 1r^ i1 2 ,
n
n
[3.85]
[3.86]
where g^ 1 5 1 g i51wi1r^ i1 2 / 1 g i51r^ 2i1 2 , as can be seen by squaring each term in (3.86), summing, and
then canceling terms. Because (3.86) is just the sum of squared residuals from the simple regression
of wi1 on r^ i1—remember that the sample average of r^ i1 is zero—(3.86) must be nonnegative. This
completes the proof.
i51
n
58860_ch03_hr_066-116.indd 116
n
10/18/18 4:23 PM
chapter
4
Multiple Regression
Analysis: Inference
T
his chapter continues our treatment of multiple regression analysis. We now turn to the problem
of testing hypotheses about the parameters in the population regression model. We begin in
Section 4-1 by finding the distributions of the OLS estimators under the added assumption that
the population error is normally distributed. Sections 4-2 and 4-3 cover hypothesis testing about individual parameters, while Section 4-4 discusses how to test a single hypothesis involving more than
one parameter. We focus on testing multiple restrictions in Section 4-5 and pay particular attention to
determining whether a group of independent variables can be omitted from a model.
4-1 Sampling Distributions of the OLS Estimators
Up to this point, we have formed a set of assumptions under which OLS is unbiased; we have also
derived and discussed the bias caused by omitted variables. In Section 3-4, we obtained the variances
of the OLS estimators under the Gauss-Markov assumptions. In Section 3-5, we showed that this variance is smallest among linear unbiased estimators.
Knowing the expected value and variance of the OLS estimators is useful for describing the
precision of the OLS estimators. However, in order to perform statistical inference, we need to know
more than just the first two moments of b^ j; we need to know the full sampling distribution of the b^ j.
Even under the Gauss-Markov assumptions, the distribution of b^ j can have virtually any shape.
When we condition on the values of the independent variables in our sample, it is clear that the
sampling distributions of the OLS estimators depend on the underlying distribution of the errors. To
make the sampling distributions of the b^ j tractable, we now assume that the unobserved error is normally distributed in the population. We call this the normality assumption.
117
58860_ch04_hr_117-162.indd 117
10/19/18 8:32 PM
118
PART 1 Regression Analysis with Cross-Sectional Data
Assumption MLR.6
Normality
The population error u is independent of the explanatory variables x1, x2, . . . , xk and is normally distributed with zero mean and variance s2: u , Normal 1 0,s2 2 .
Assumption MLR.6 is much stronger than any of our previous assumptions. In
fact, because u is independent of the xj under MLR.6, E 1 u 0 x1, . . . , xk 2 5 E 1 u 2 5 0 and
Var 1 u 0 x1, . . . , xk 2 5 Var 1 u 2 5 s2. Thus, if we make Assumption MLR.6, then we are necessarily
assuming MLR.4 and MLR.5. To emphasize that we are assuming more than before, we will refer to
the full set of Assumptions MLR.1 through MLR.6.
For cross-sectional regression applications, Assumptions MLR.1 through MLR.6 are called
the classical linear model (CLM) assumptions. Thus, we will refer to the model under these six
assumptions as the classical linear model. It is best to think of the CLM assumptions as containing
all of the Gauss-Markov assumptions plus the assumption of a normally distributed error term.
Under the CLM assumptions, the OLS estimators b^ 0, b^ 1, . . . , b^ k have a stronger efficiency property than they would under the Gauss-Markov assumptions. It can be shown that the OLS estimators are
the minimum variance unbiased estimators, which means that OLS has the smallest variance among
unbiased estimators; we no longer have to restrict our comparison to estimators that are linear in the yi.
This property of OLS under the CLM assumptions is discussed further in Advanced Treatment E.
A succinct way to summarize the population assumptions of the CLM is
y 0 x , Normal 1 b0 1 b1x1 1 b2x2 1 p 1 bkxk,s2 2 ,
where x is again shorthand for 1 x1, . . . , xk 2 . Thus, conditional on x, y has a normal distribution with
mean linear in x1, . . . , xk and a constant variance. For a single independent variable x, this situation is
shown in Figure 4.1.
Figure 4.1 The homoskedastic normal distribution with a single explanatory variable.
f(ylx)
y
normal distributions
x1
E(ylx) 5
x2
0
1
x
1
x3
x
58860_ch04_hr_117-162.indd 118
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
119
The argument justifying the normal distribution for the errors usually runs something like this:
Because u is the sum of many different unobserved factors affecting y, we can invoke the central limit
theorem (CLT) (see Math Refresher C) to conclude that u has an approximate normal distribution.
This argument has some merit, but it is not without weaknesses. First, the factors in u can have very
different distributions in the population (for example, ability and quality of schooling in the error in a
wage equation). Although the CLT can still hold in such cases, the normal approximation can be poor
depending on how many factors appear in u and how different their distributions are.
A more serious problem with the CLT argument is that it assumes that all unobserved factors
affect y in a separate, additive fashion. Nothing guarantees that this is so. If u is a complicated function of the unobserved factors, then the CLT argument does not really apply.
In any application, whether normality of u can be assumed is really an empirical matter. For
example, there is no theorem that says wage conditional on educ, exper, and tenure is normally distributed. If anything, simple reasoning suggests that the opposite is true: because wage can never
be less than zero, it cannot, strictly speaking, have a normal distribution. Further, because there are
minimum wage laws, some fraction of the population earns exactly the minimum wage, which also
violates the normality assumption. Nevertheless, as a practical matter, we can ask whether the conditional wage distribution is “close” to being normal. Past empirical evidence suggests that normality is
not a good assumption for wages.
Often, using a transformation, especially taking the log, yields a distribution that is closer to
normal. For example, something like log(price) tends to have a distribution that looks more normal
than the distribution of price. Again, this is an empirical issue. We will discuss the consequences of
nonnormality for statistical inference in Chapter 5.
There are some applications where MLR.6 is clearly false, as can be demonstrated with simple
introspection. Whenever y takes on just a few values it cannot have anything close to a normal distribution. The dependent variable in Example 3.5 provides a good example. The variable narr86, the
number of times a young man was arrested in 1986, takes on a small range of integer values and is
zero for most men. Thus, narr86 is far from being normally distributed. What can be done in these
cases? As we will see in Chapter 5—and this is important—nonnormality of the errors is not a serious
problem with large sample sizes. For now, we just make the normality assumption.
Normality of the error term translates into normal sampling distributions of the OLS estimators:
Theorem
4.1
Normal Sampling Distributions
Under the CLM assumptions MLR.1 through MLR.6, conditional on the sample values of the independent variables,
b^ j , Normal 3 bj,Var 1 b^ j 2 4 ,
[4.1]
where Var 1 b^ j 2 was given in Chapter 3 [equation (3.51)]. Therefore,
1 b^ j 2 bj 2 /sd 1 b^ j 2 , Normal 1 0,1 2 .
The proof of (4.1) is not that difficult, given the properties of normally distributed random variables in
n
Math Refresher B. Each b^ j can be written as b^ j 5 bj 1 g i51wijui, where wij 5 r^ ij/SSRj, r^ ij is the ith residual from the regression of the xj on all the other independent variables, and SSRj is the sum of squared
residuals from this regression [see equation (3.65)]. Because the wij depend only on the independent variables, they can be treated as nonrandom. Thus, b^ j is just a linear combination of the errors in the sample
5 ui: i 5 1, 2, . . . , n 6 . Under Assumption MLR.6 (and the random sampling Assumption MLR.2), the
errors are independent, identically distributed Normal 1 0,s2 2 random variables. An important fact about
independent normal random variables is that a linear combination of such random variables is normally
58860_ch04_hr_117-162.indd 119
10/19/18 8:32 PM
120
PART 1 Regression Analysis with Cross-Sectional Data
distributed (see Math Refresher B). This basically
completes the proof. In Section 3-3, we showed that
Suppose that u is independent of the
E 1 b^ j 2 5 bj, and we derived Var 1 b^ j 2 in Section 3-4;
explanatory variables, and it takes on the
there is no need to re-derive these facts.
values 22, 21, 0, 1, and 2 with equal probThe second part of this theorem follows immeability of 1/5. Does this violate the Gaussdiately from the fact that when we standardize a norMarkov assumptions? Does this violate the
mal random variable by subtracting off its mean and
CLM assumptions?
dividing by its standard deviation, we end up with a
standard normal random variable.
The conclusions of Theorem 4.1 can be strengthened. In addition to (4.1), any linear combination of the b^ 0, b^ 1, . . . , b^ k is also normally distributed, and any subset of the b^ j has a joint normal
distribution. These facts underlie the testing results in the remainder of this chapter. In Chapter 5, we
will show that the normality of the OLS estimators is still approximately true in large samples even
without normality of the errors.
G o i n g F u rt h e r 4 . 1
4-2 Testing Hypotheses about a Single Population Parameter: The t Test
This section covers the very important topic of testing hypotheses about any single parameter in the
population regression function. The population model can be written as
y 5 b0 1 b1x1 1 p 1 bkxk 1 u,
[4.2]
and we assume that it satisfies the CLM assumptions. We know that OLS produces unbiased estimators of the bj. In this section, we study how to test hypotheses about a particular bj. For a full understanding of hypothesis testing, one must remember that the bj are unknown features of the population,
and we will never know them with certainty. Nevertheless, we can hypothesize about the value of bj
and then use statistical inference to test our hypothesis.
In order to construct hypotheses tests, we need the following result:
Theorem
4.2
t Distribution for the Standardized Estimators
Under the CLM assumptions MLR.1 through MLR.6,
1 b^ j 2 bj 2 /se 1 b^ j 2 , tn2k21 5 tdf,
[4.3]
where k 1 1 is the number of unknown parameters in the population model y 5 b0 1 b1x1 1 p 1
bk xk 1 u (k slope parameters and the intercept b0) and n 2 k 2 1 is the degrees of freedom (df).
This result differs from Theorem 4.1 in some notable respects. Theorem 4.1 showed that, under the
CLM assumptions, 1 b^ j 2 bj 2 /sd 1 b^ j 2 , Normal 1 0,1 2 . The t distribution in (4.3) comes from the
fact that the constant s in sd 1 b^ j 2 has been replaced with the random variable s^ . The proof that this
leads to a t distribution with n 2 k 2 1 degrees of freedom is difficult and not especially instructive.
Essentially, the proof shows that (4.3) can be written as the ratio of the standard normal random variable 1 b^ j 2 bj 2 /sd 1 b^ j 2 over the square root of s^ 2 /s2. These random variables can be shown to be independent, and 1 n 2 k 2 1 2 s^ 2 /s2 , x2n2k21. The result then follows from the definition of a t random
variable (see Section B-5 in Math Refresher B).
Theorem 4.2 is important in that it allows us to test hypotheses involving the bj. In most applications, our primary interest lies in testing the null hypothesis
H0: bj 5 0,
58860_ch04_hr_117-162.indd 120
[4.4]
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
121
where j corresponds to any of the k independent variables. It is important to understand what (4.4)
means and to be able to describe this hypothesis in simple language for a particular application.
Because bj measures the partial effect of xj on (the expected value of) y, after controlling for all other
independent variables, (4.4) means that, once x1, x2, . . . , xj21, xj11, . . . , xk have been accounted for,
xj has no effect on the expected value of y. We cannot state the null hypothesis as “xj does have a partial effect on y” because this is true for any value of bj other than zero. Classical testing is suited for
testing simple hypotheses like (4.4).
As an example, consider the wage equation
log 1 wage 2 5 b0 1 b1educ 1 b2exper 1 b3tenure 1 u.
The null hypothesis H0: b2 5 0 means that, once education and tenure have been accounted for, the
number of years in the workforce (exper) has no effect on hourly wage. This is an economically interesting hypothesis. If it is true, it implies that a person’s work history prior to the current employment
does not affect wage. If b2 . 0, then prior work experience contributes to productivity, and hence
to wage.
You probably remember from your statistics course the rudiments of hypothesis testing for the
mean from a normal population. (This is reviewed in Math Refresher C.) The mechanics of testing (4.4)
in the multiple regression context are very similar. The hard part is obtaining the coefficient estimates,
the standard errors, and the critical values, but most of this work is done automatically by econometrics
software. Our job is to learn how regression output can be used to test hypotheses of interest.
The statistic we use to test (4.4) (against any alternative) is called “the” t statistic or “the” t ratio
of b^ j and is defined as
tb^ j ; b^ j/se 1 b^ j 2 .
[4.5]
We have put “the” in quotation marks because, as we will see shortly, a more general form of the
t statistic is needed for testing other hypotheses about bj. For now, it is important to know that (4.5)
is suitable only for testing (4.4). For particular applications, it is helpful to index t statistics using the
name of the independent variable; for example, teduc would be the t statistic for b^ educ.
The t statistic for b^ j is simple to compute given b^ j and its standard error. In fact, most regression
packages do the division for you and report the t statistic along with each coefficient and its standard
error.
Before discussing how to use (4.5) formally to test H0: bj 5 0, it is useful to see why tb^ j has
features that make it reasonable as a test statistic to detect bj 2 0. First, because se 1 b^ j 2 is always
­positive, tb^ j has the same sign as b^ j: if b^ j is positive, then so is tb^ j and if b^ j is negative, so is tb^ j. Second,
for a given value of se 1 b^ j 2 , a larger value of b^ j leads to larger values of tb^ j. If b^ j becomes more negative, so does tb^ j.
Because we are testing H0: bj 5 0, it is only natural to look at our unbiased estimator of bj, b^ j,
for guidance. In any interesting application, the point estimate b^ j will never exactly be zero, whether
or not H0 is true. The question is: How far is b^ j from zero? A sample value of b^ j very far from zero
provides evidence against H0: bj 5 0. However, we must recognize that there is a sampling error in
our estimate b^ j, so the size of b^ j must be weighed against its sampling error. Because the standard
error of b^ j is an estimate of the standard deviation of b^ j, tb^ j measures how many estimated standard
deviations b^ j is away from zero. This is precisely what we do in testing whether the mean of a population is zero, using the standard t statistic from introductory statistics. Values of tb^ j sufficiently far from
zero will result in a rejection of H0. The precise rejection rule depends on the alternative hypothesis
and the chosen significance level of the test.
Determining a rule for rejecting (4.4) at a given significance level—that is, the probability of
rejecting H0 when it is true—requires knowing the sampling distribution of tb^ j when H0 is true. From
Theorem 4.2, we know this to be tn2k21. This is the key theoretical result needed for testing (4.4).
Before proceeding, it is important to remember that we are testing hypotheses about the population parameters. We are not testing hypotheses about the estimates from a particular sample. Thus, it
58860_ch04_hr_117-162.indd 121
10/19/18 8:32 PM
122
PART 1 Regression Analysis with Cross-Sectional Data
never makes sense to state a null hypothesis as “H0: b^ 1 5 0” or, even worse, as “H0: .237 5 0” when
the estimate of a parameter is .237 in the sample. We are testing whether the unknown population
value, b1, is zero.
Some treatments of regression analysis define the t statistic as the absolute value of (4.5), so that
the t statistic is always positive. This practice has the drawback of making testing against one-sided
alternatives clumsy. Throughout this text, the t statistic always has the same sign as the corresponding
OLS coefficient estimate.
4-2a Testing against One-Sided Alternatives
To determine a rule for rejecting H0, we need to decide on the relevant alternative hypothesis. First,
consider a one-sided alternative of the form
H1: bj . 0.
[4.6]
When we state the alternative as in equation (4.6), we are really saying that the null hypothesis is
H0: bj # 0. For example, if bj is the coefficient on education in a wage regression, we only care about
detecting that bj is different from zero when bj is actually positive. You may remember from introductory statistics that the null value that is hardest to reject in favor of (4.6) is bj 5 0. In other words, if we
reject the null bj 5 0 then we automatically reject bj , 0. Therefore, it suffices to act as if we are testing
H0: bj 5 0 against H1: bj . 0, effectively ignoring bj , 0, and that is the approach we take in this book.
How should we choose a rejection rule? We must first decide on a significance level (“level”
for short) or the probability of rejecting H0 when it is in fact true. For concreteness, suppose we have
decided on a 5% significance level, as this is the most popular choice. Thus, we are willing to mistakenly reject H0 when it is true 5% of the time. Now, while tb^ j has a t distribution under H0—so that it
has zero mean—under the alternative bj . 0, the expected value of tb^ j is positive. Thus, we are looking for a “sufficiently large” positive value of tb^ j in order to reject H0: bj 5 0 in favor of H1: bj . 0.
Negative values of tb^ j provide no evidence in favor of H1.
The definition of “sufficiently large,” with a 5% significance level, is the 95th percentile in a
t distribution with n 2 k 2 1 degrees of freedom; denote this by c. In other words, the rejection rule
is that H0 is rejected in favor of H1 at the 5% significance level if
tb^ j . c.
[4.7]
By our choice of the critical value, c, rejection of H0 will occur for 5% of all random samples when
H0 is true.
The rejection rule in (4.7) is an example of a one-tailed test. To obtain c, we only need the significance level and the degrees of freedom. For example, for a 5% level test and with n 2 k 2 1 5 28
degrees of freedom, the critical value is c 5 1.701. If tb^ j # 1.701, then we fail to reject H0 in favor of
(4.6) at the 5% level. Note that a negative value for tb^ j, no matter how large in absolute value, leads to
a failure in rejecting H0 in favor of (4.6). (See Figure 4.2.)
The same procedure can be used with other significance levels. For a 10% level test and if
df 5 21, the critical value is c 5 1.323. For a 1% significance level and if df 5 21, c 5 2.518. All
of these critical values are obtained directly from Table G.2. You should note a pattern in the critical
values: as the significance level falls, the critical value increases, so that we require a larger and larger
value of tb^ j in order to reject H0. Thus, if H0 is rejected at, say, the 5% level, then it is automatically
rejected at the 10% level as well. It makes no sense to reject the null hypothesis at, say, the 5% level
and then to redo the test to determine the outcome at the 10% level.
As the degrees of freedom in the t distribution get large, the t distribution approaches the standard
normal distribution. For example, when n 2 k 2 1 5 120, the 5% critical value for the one-sided
alternative (4.7) is 1.658, compared with the standard normal value of 1.645. These are close enough
for practical purposes; for degrees of freedom greater than 120, one can use the standard normal
­critical values.
58860_ch04_hr_117-162.indd 122
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
123
Figure 4.2 5% rejection rule for the alternative H1: bj . 0 with 28 df.
area = .05
0
1.701
Example 4.1
rejection
region
Hourly Wage Equation
Using the data in WAGE1 gives the estimated equation
log 1 wage 2 5 .284 1 .092 educ 1 .0041 exper 1 .022 tenure
1 .104 2 1 .007 2
1 .0017 2
1 .003 2
n 5 526, R2 5 .316,
where standard errors appear in parentheses below the estimated coefficients. We will follow this
convention throughout the text. This equation can be used to test whether the return to exper, controlling for educ and tenure, is zero in the population, against the alternative that it is positive. Write this
as H0: bexper 5 0 versus H1: bexper . 0. (In applications, indexing a parameter by its associated variable name is a nice way to label parameters, because the numerical indices that we use in the general
model are arbitrary and can cause confusion.) Remember that bexper denotes the unknown population
parameter. It is nonsense to write “H0: .0041 5 0” or “H0: b^ exper 5 0.”
Because we have 522 degrees of freedom, we can use the standard normal critical values. The 5%
critical value is 1.645, and the 1% critical value is 2.326. The t statistic for b^ exper is
texper 5 .0041/.0017 < 2.41,
and so b^ exper, or exper, is statistically significant even at the 1% level. We also say that “b^ exper is statistically greater than zero at the 1% significance level.”
The estimated return for another year of experience, holding tenure and education fixed, is not
especially large. For example, adding three more years increases log(wage) by 3 1 .0041 2 5 .0123, so
wage is only about 1.2% higher. Nevertheless, we have persuasively shown that the partial effect of
experience is positive in the population.
58860_ch04_hr_117-162.indd 123
10/19/18 8:32 PM
124
PART 1 Regression Analysis with Cross-Sectional Data
The one-sided alternative that the parameter is less than zero,
[4.8]
H1: bj , 0,
G o i n g F u rt h e r 4 . 2
Let community loan approval rates be determined by
apprate 5 b0 1 b1percmin 1 b2avginc
1 b3avgwlth 1 b4avgdebt 1 u,
where percmin is the percentage minority in
the community, avginc is average income,
avgwlth is average wealth, and avgdebt is
some measure of average debt obligations.
how do you state the null hypothesis that
there is no difference in loan rates across
neighborhoods due to racial and ethnic
composition, when average income, average wealth, and average debt have been
controlled for? how do you state the alternative that there is discrimination against
minorities in loan approval rates?
also arises in applications. The rejection rule for
alternative (4.8) is just the mirror image of the previous case. Now, the critical value comes from the left
tail of the t distribution. In practice, it is easiest to
think of the rejection rule as
tb^ j , 2c,
[4.9]
where c is the critical value for the alternative
H1: bj . 0. For simplicity, we always assume c
is positive, because this is how critical values are
reported in t tables, and so the critical value 2c is a
negative number.
For example, if the significance level is 5% and
the degrees of freedom is 18, then c 5 1.734, and so
H0: bj 5 0 is rejected in favor of H1: bj , 0 at the
5% level if tb^ j , 21.734. It is important to remember
that, to reject H0 against the negative alternative (4.8),
we must get a negative t statistic. A positive t ratio,
no matter how large, provides no evidence in favor of
(4.8). The rejection rule is illustrated in Figure 4.3.
Figure 4.3 5% rejection rule for the alternative H1: bj , 0 with 18 df.
area = .05
0
rejection
region
58860_ch04_hr_117-162.indd 124
–1.734
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
Example 4.2
125
Student Performance and School Size
There is much interest in the effect of school size on student performance. (See, for example, The
New York Times Magazine, 5/28/95.) One claim is that, everything else being equal, students at
smaller schools fare better than those at larger schools. This hypothesis is assumed to be true even
after accounting for differences in class sizes across schools.
The file MEAP93 contains data on 408 high schools in Michigan for the year 1993. We can use
these data to test the null hypothesis that school size has no effect on standardized test scores against
the alternative that size has a negative effect. Performance is measured by the percentage of students
receiving a passing score on the Michigan Educational Assessment Program (MEAP) standardized
tenth-grade math test (math10). School size is measured by student enrollment (enroll). The null
hypothesis is H0: benroll 5 0, and the alternative is H1: benroll , 0. For now, we will control for two
other factors, average annual teacher compensation (totcomp) and the number of staff per one thousand students (staff). Teacher compensation is a measure of teacher quality, and staff size is a rough
measure of how much attention students receive.
The estimated equation, with standard errors in parentheses, is
math10 5 2.274 1 .00046 totcomp 1 .048 staff 2 .00020 enroll
1 6.113 2 1 .00010 2
1 .040 2
1 .00022 2
n 5 408, R2 5 .0541.
The coefficient on enroll, 2.00020, is in accordance with the conjecture that larger schools hamper
performance: higher enrollment leads to a lower percentage of students with a passing tenth-grade
math score. (The coefficients on totcomp and staff also have the signs we expect.) The fact that enroll
has an estimated coefficient different from zero could just be due to sampling error; to be convinced
of an effect, we need to conduct a t test.
Because n 2 k 2 1 5 408 2 4 5 404, we use the standard normal critical value. At the 5% level,
the critical value is 21.65; the t statistic on enroll must be less than 21.65 to reject H0 at the 5% level.
The t statistic on enroll is 2.00020/.00022 < 2.91, which is larger than 21.65: we fail to reject
H0 in favor of H1 at the 5% level. In fact, the 15% critical value is 21.04, and because 2.91 . 21.04,
we fail to reject H0 even at the 15% level. We conclude that enroll is not statistically significant at the
15% level.
The variable totcomp is statistically significant even at the 1% significance level because its
t statistic is 4.6. On the other hand, the t statistic for staff is 1.2, and so we cannot reject H0: bstaff 5 0
against H1: bstaff . 0 even at the 10% significance level. (The critical value is c 5 1.28 from the standard normal distribution.)
To illustrate how changing functional form can affect our conclusions, we also estimate the
model with all independent variables in logarithmic form. This allows, for example, the school size
effect to diminish as school size increases. The estimated equation is
math10 5 2207.66 1 21.16 log 1 totcomp 2 1 3.98 log 1 staff 2 2 1.29 log 1 enroll 2
1 48.70 2
1 4.06 2
1 4.19 2
1 0.69 2
n 5 408, R2 5 .0654.
The t statistic on log(enroll) is about 21.87; because this is below the 5% critical value 21.65, we
reject H0: blog1enroll2 5 0 in favor of H1: blog1enroll2 , 0 at the 5% level.
In Chapter 2, we encountered a model in which the dependent variable appeared in its original form (called level form), while the independent variable appeared in log form (called level-log
model). The interpretation of the parameters is the same in the multiple regression context, except,
of course, that we can give the parameters a ceteris paribus interpretation. Holding totcomp and staff
fixed, we have Dmath10 5 21.29 3 Dlog 1 enroll 2 4 , so that
Dmath10 < 2 1 1.29/100 2 1 %Denroll 2 < 2.013 1 %Denroll 2 .
58860_ch04_hr_117-162.indd 125
10/19/18 8:32 PM
126
PART 1 Regression Analysis with Cross-Sectional Data
Once again, we have used the fact that the change in log(enroll), when multiplied by 100, is approximately the percentage change in enroll. Thus, if enrollment is 10% higher at a school, math10 is predicted to be .013 1 10 2 5 0.13 percentage points lower (math10 is measured as a percentage).
Which model do we prefer, the one using the level of enroll or the one using log(enroll)? In the
level-level model, enrollment does not have a statistically significant effect, but in the level-log model
it does. This translates into a higher R-squared for the level-log model, which means we explain more
of the variation in math10 by using enroll in logarithmic form (6.5% to 5.4%). The level-log model
is preferred because it more closely captures the relationship between math10 and enroll. We will say
more about using R-squared to choose functional form in Chapter 6.
4-2b Two-Sided Alternatives
In applications, it is common to test the null hypothesis H0: bj 5 0 against a two-sided alternative;
that is,
H1: bj 2 0.
[4.10]
Under this alternative, xj has a ceteris paribus effect on y without specifying whether the effect is positive or negative. This is the relevant alternative when the sign of bj is not well determined by theory
(or common sense). Even when we know whether bj is positive or negative under the alternative, a
two-sided test is often prudent. At a minimum, using a two-sided alternative prevents us from looking at the estimated equation and then basing the alternative on whether b^ j is positive or negative.
Using the regression estimates to help us formulate the null or alternative hypotheses is not allowed
Figure 4.4 5% rejection rule for the alternative H1: bj 2 0 with 25 df.
area = .025
area = .025
0
rejection
region
58860_ch04_hr_117-162.indd 126
–2.06
2.06
rejection
region
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
127
because classical statistical inference presumes that we state the null and alternative about the population before looking at the data. For example, we should not first estimate the equation relating math
performance to enrollment, note that the estimated effect is negative, and then decide the relevant
alternative is H1: benroll , 0.
When the alternative is two-sided, we are interested in the absolute value of the t statistic. The
rejection rule for H0: bj 5 0 against (4.10) is
0 tb^ j 0 . c,
[4.11]
where 0 # 0 denotes absolute value and c is an appropriately chosen critical value. To find c, we again
specify a significance level, say 5%. For a two-tailed test, c is chosen to make the area in each tail
of the t distribution an equal 2.5%. In other words, c is the 97.5th percentile in the t distribution with
n 2 k 2 1 degrees of freedom. When n 2 k 2 1 5 25, the 5% critical value for a two-sided test is
c 5 2.060. Figure 4.4 provides an illustration of this distribution.
When a specific alternative is not stated, it is usually considered to be two-sided. In the remainder
of this text, the default will be a two-sided alternative, and 5% will be the default significance level.
When carrying out empirical econometric analysis, it is always a good idea to be explicit about the
alternative and the significance level. If H0 is rejected in favor of (4.10) at the 5% level, we usually
say that “xj is statistically significant, or statistically different from zero, at the 5% level.” If H0 is not
rejected, we say that “xj is statistically insignificant at the 5% level.”
Example 4.3
Determinants of College GPA
We use the data in GPA1 to estimate a model explaining college GPA (colGPA), with the average
number of lectures missed per week (skipped) as an additional explanatory variable. The estimated
model is
colGPA 5 1.39 1 .412 hsGPA 1 .015 ACT 2 .083 skipped
1 .33 2 1 .094 2
1 .011 2
1 .026 2
n 5 141, R2 5 .234.
We can easily compute t statistics to see which variables are statistically significant, using a twosided alternative in each case. The 5% critical value is about 1.96, because the degrees of freedom
1 141 2 4 5 137 2 is large enough to use the standard normal approximation. The 1% critical value is
about 2.58.
The t statistic on hsGPA is 4.38, which is significant at very small significance levels. Thus, we
say that “hsGPA is statistically significant at any conventional significance level.” The t statistic on
ACT is 1.36, which is not statistically significant at the 10% level against a two-sided alternative. The
coefficient on ACT is also practically small: a 10-point increase in ACT, which is large, is predicted
to increase colGPA by only .15 points. Thus, the variable ACT is practically, as well as statistically,
insignificant.
The coefficient on skipped has a t statistic of 2.083/.026 5 23.19, so skipped is statistically
significant at the 1% significance level 1 3.19 . 2.58 2 . This coefficient means that another lecture
missed per week lowers predicted colGPA by about .083. Thus, holding hsGPA and ACT fixed, the
predicted difference in colGPA between a student who misses no lectures per week and a student who
misses five lectures per week is about .42. Remember that this says nothing about specific students;
rather, .42 is the estimated average across a subpopulation of students.
In this example, for each variable in the model, we could argue that a one-sided alternative is
appropriate. The variables hsGPA and skipped are very significant using a two-tailed test and have the
signs that we expect, so there is no reason to do a one-tailed test. On the other hand, against a onesided alternative 1 b3 . 0 2 , ACT is significant at the 10% level but not at the 5% level. This does not
change the fact that the coefficient on ACT is pretty small.
58860_ch04_hr_117-162.indd 127
10/19/18 8:32 PM
128
PART 1 Regression Analysis with Cross-Sectional Data
4-2c Testing Other Hypotheses about bj
Although H0: bj 5 0 is the most common hypothesis, we sometimes want to test whether bj is equal
to some other given constant. Two common examples are bj 5 1 and bj 5 21. Generally, if the null
is stated as
[4.12]
H0: bj 5 aj,
where aj is our hypothesized value of bj, then the appropriate t statistic is
t 5 1 b^ j 2 aj 2 /se 1 b^ j 2 .
As before, t measures how many estimated standard deviations b^ j is away from the hypothesized
value of bj. The general t statistic is usefully written as
t5
1 estimate 2 hypothesized value 2
standard error
.
[4.13]
Under (4.12), this t statistic is distributed as tn2k21 from Theorem 4.2. The usual t statistic is obtained
when aj 5 0.
We can use the general t statistic to test against one-sided or two-sided alternatives. For example,
if the null and alternative hypotheses are H0: bj 5 1 and H1: bj . 1, then we find the critical value
for a one-sided alternative exactly as before: the difference is in how we compute the t statistic, not in
how we obtain the appropriate c. We reject H0 in favor of H1 if t . c. In this case, we would say that
“b^ j is statistically greater than one” at the appropriate significance level.
Example 4.4
Campus Crime and Enrollment
Consider a simple model relating the annual number of crimes on college campuses (crime) to student
enrollment (enroll):
log 1 crime 2 5 b0 1 b1log 1 enroll 2 1 u.
This is a constant elasticity model, where b1 is the elasticity of crime with respect to enrollment. It
is not much use to test H0: b1 5 0, as we expect the total number of crimes to increase as the size
of the campus increases. A more interesting hypothesis to test would be that the elasticity of crime
with respect to enrollment is one: H0: b1 5 1. This means that a 1% increase in enrollment leads to,
on average, a 1% increase in crime. A noteworthy alternative is H1: b1 . 1, which implies that a 1%
increase in enrollment increases campus crime by more than 1%. If b1 . 1, then, in a relative sense—
not just an absolute sense—crime is more of a problem on larger campuses. One way to see this is to
take the exponential of the equation:
crime 5 exp 1 b0 2 enrollb1exp 1 u 2 .
(See Math Refresher A for properties of the natural logarithm and exponential functions.) For b0 5 0
and u 5 0, this equation is graphed in Figure 4.5 for b1 , 1, b1 5 1, and b1 . 1.
We test b1 5 1 against b1 . 1 using data on 97 colleges and universities in the United States for
the year 1992, contained in the data file CAMPUS. The data come from the FBI’s Uniform Crime
Reports, and the average number of campus crimes in the sample is about 394, while the average
enrollment is about 16,076. The estimated equation (with estimates and standard errors rounded to
two decimal places) is
log 1 crime 2 5 26.63 1 1.27 log 1 enroll 2
1 1.03 2 1 0.11 2
[4.14]
2
n 5 97, R 5 .585.
58860_ch04_hr_117-162.indd 128
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
129
Figure 4.5 Graph of crime 5 enroll b1 for b1 , 1, b1 5 1, and b1 . 1.
crime
1
1
=1
>1
1
<1
0
0
enroll
The estimated elasticity of crime with respect to enroll, 1.27, is in the direction of the alternative b1 . 1. But is there enough evidence to conclude that b1 . 1? We need to be careful in
testing this hypothesis, especially because the statistical output of standard regression packages
is much more complex than the simplified output reported in equation (4.14). Our first instinct
might be to construct “the” t statistic by taking the coefficient on log(enroll) and dividing it
by its standard error, which is the t statistic reported by a regression package. But this is the
wrong statistic for testing H0: b1 5 1. The correct t statistic is obtained from (4.13): we subtract
the hypothesized value, unity, from the estimate and divide the result by the standard error of
b^ 1: t 5 1 1.27 2 1 2 /.11 5 .27/.11 < 2.45. The one-sided 5% critical value for a t distribution with
97 2 2 5 95 df is about 1.66 (using df 5 120 2 , so we clearly reject b1 5 1 in favor of b1 . 1 at
the 5% level. In fact, the 1% critical value is about 2.37, and so we reject the null in favor of the
alternative at even the 1% level.
We should keep in mind that this analysis holds no other factors constant, so the elasticity of
1.27 is not necessarily a good estimate of ceteris paribus effect. It could be that larger enrollments are
correlated with other factors that cause higher crime: larger schools might be located in higher crime
areas. We could control for this by collecting data on crime rates in the local city.
For a two-sided alternative, for example H0: bj 5 21, H1: b1 2 21, we still compute the t statistic as in (4.13): t 5 1 b^ j 1 1 2 /se 1 b^ j 2 (notice how subtracting 21 means adding 1). The rejection rule
is the usual one for a two-sided test: reject H0 if 0 t 0 . c, where c is a two-tailed critical value. If H0
is rejected, we say that “b^ j is statistically different from negative one” at the appropriate significance
level.
58860_ch04_hr_117-162.indd 129
10/19/18 8:32 PM
130
PART 1 Regression Analysis with Cross-Sectional Data
Example 4.5
Housing Prices and Air Pollution
For a sample of 506 communities in the Boston area, we estimate a model relating median housing
price (price) in the community to various community characteristics: nox is the amount of nitrogen
oxide in the air, in parts per million; dist is a weighted distance of the community from five employment centers, in miles; rooms is the average number of rooms in houses in the community; and stratio
is the average student-teacher ratio of schools in the community. The population model is
log 1 price 2 5 b0 1 b1log 1 nox 2 1 b2log 1 dist 2 1 b3rooms 1 b4stratio 1 u.
Thus, b1 is the elasticity of price with respect to nox. We wish to test H0: b1 5 21 against the alternative H1: b1 2 21. The t statistic for doing this test is t 5 1 b^ 1 1 1 2 /se 1 b^ 1 2 .
Using the data in HPRICE2, the estimated model is
log 1 price 2 5 11.08 2 .954 log 1 nox 2 2 .134 log 1 dist 2 1 .255 rooms 2 .052 stratio
1 0.32 2 1 .117 2
1 .043 2
1 .019 2
1 .006 2
n 5 506, R2 5 .581.
The slope estimates all have the anticipated signs. Each coefficient is statistically different from
zero at very small significance levels, including the coefficient on log(nox). But we do not want
to test that b1 5 0. The null hypothesis of interest is H0: b1 5 21, with corresponding t statistic
1 2.954 1 1 2 /.117 5 .393. There is little need to look in the t table for a critical value when the
t statistic is this small: the estimated elasticity is not statistically different from 21 even at very large
significance levels. Controlling for the factors we have included, there is little evidence that the elasticity is different from 21.
4-2d Computing p-Values for t Tests
So far, we have talked about how to test hypotheses using a classical approach: after stating the alternative hypothesis, we choose a significance level, which then determines a critical value. Once the
critical value has been identified, the value of the t statistic is compared with the critical value, and the
null is either rejected or not rejected at the given significance level.
Even after deciding on the appropriate alternative, there is a component of arbitrariness to the
classical approach, which results from having to choose a significance level ahead of time. Different
researchers prefer different significance levels, depending on the particular application. There is no
“correct” significance level.
Committing to a significance level ahead of time can hide useful information about the outcome of
a hypothesis test. For example, suppose that we wish to test the null hypothesis that a parameter is zero
against a two-sided alternative, and with 40 degrees of freedom we obtain a t statistic equal to 1.85.
The null hypothesis is not rejected at the 5% level, because the t statistic is less than the two-tailed
critical value of c 5 2.021. A researcher whose agenda is not to reject the null could simply report this
outcome along with the estimate: the null hypothesis is not rejected at the 5% level. Of course, if the
t statistic, or the coefficient and its standard error, are reported, then we can also determine that the null
hypothesis would be rejected at the 10% level, because the 10% critical value is c 5 1.684.
Rather than testing at different significance levels, it is more informative to answer the following
question: given the observed value of the t statistic, what is the smallest significance level at which
the null hypothesis would be rejected? This level is known as the p-value for the test (see Math
Refresher C). In the previous example, we know the p-value is greater than .05, because the null is
not rejected at the 5% level, and we know that the p-value is less than .10, because the null is rejected
at the 10% level. We obtain the actual p-value by computing the probability that a t random variable,
with 40 df, is larger than 1.85 in absolute value. That is, the p-value is the significance level of the test
when we use the value of the test statistic, 1.85 in the above example, as the critical value for the test.
This p-value is shown in Figure 4.6.
58860_ch04_hr_117-162.indd 130
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
131
Figure 4.6 Obtaining the p-value against a two-sided alternative, when t 5 1.85 and df 5 40.
area = .9282
area = .0359
area = .0359
–1.85
0
1.85
Because a p-value is a probability, its value is always between zero and one. In order to compute
p-values, we either need extremely detailed printed tables of the t distribution—which is not very
practical—or a computer program that computes areas under the probability density function of the
t distribution. Most modern regression packages have this capability. Some packages compute p-values
routinely with each OLS regression, but only for certain hypotheses. If a regression package reports
a p-value along with the standard OLS output, it is almost certainly the p-value for testing the null
hypothesis H0: bj 5 0 against the two-sided alternative. The p-value in this case is
P1 0T0 . 0t0 2,
[4.15]
where, for clarity, we let T denote a t distributed random variable with n 2 k 2 1 degrees of freedom
and let t denote the numerical value of the test statistic.
The p-value nicely summarizes the strength or weakness of the empirical evidence against the
null hypothesis. Perhaps its most useful interpretation is the following: the p-value is the probability
of observing a t statistic as extreme as we did if the null hypothesis is true. This means that small
p-values are evidence against the null; large p-values provide little evidence against H0. For example, if the p-value 5 .50 (reported always as a decimal, not a percentage), then we would observe a
value of the t statistic as extreme as we did in 50% of all random samples when the null hypothesis is
true; this is pretty weak evidence against H0.
In the example with df 5 40 and t 5 1.85, the p-value is computed as
p-value 5 P 1 0 T 0 . 1.85 2 5 2P 1 T . 1.85 2 5 2 1 .0359 2 5 .0718,
where P 1 T . 1.85 2 is the area to the right of 1.85 in a t distribution with 40 df. (This value was computed using the econometrics package Stata; it is not available in Table G.2.) This means that, if the
58860_ch04_hr_117-162.indd 131
10/19/18 8:32 PM
132
PART 1 Regression Analysis with Cross-Sectional Data
null hypothesis is true, we would observe an absolute value of the t statistic as large as 1.85 about
7.2 percent of the time. This provides some evidence against the null hypothesis, but we would not
reject the null at the 5% significance level.
The previous example illustrates that once the p-value has been computed, a classical test can be
carried out at any desired level. If a denotes the significance level of the test (in decimal form), then
H0 is rejected if p-value , a; otherwise, H0 is not rejected at the 100#a% level.
Computing p-values for one-sided alternatives is also quite simple. Suppose, for example, that we
test H0: bj 5 0 against H1: bj . 0. If b^ j , 0, then computing a p-value is not important: we know that
the p-value is greater than .50, which will never cause us to reject H0 in favor of H1. If b^ j . 0, then
t . 0 and the p-value is just the probability that a random t variable with the appropriate df exceeds
the value t. Some regression packages only compute p-values for two-sided alternatives. But it is
simple to obtain the one-sided p-value: just divide the two-sided p-value by 2.
If the alternative is H1: bj , 0, it makes sense to compute a p-value if b^ j , 0 (and hence t , 0):
p-value 5 P 1 T , t 2 5 P 1 T . 0 t 0 2 because the t distribution is symmetric about zero. Again, this can
be obtained as one-half of the p-value for the two-tailed test.
Because you will quickly become familiar with
the
magnitudes
of t statistics that lead to statistical
G o i n g F u rt h e r 4 . 3
significance, especially for large sample sizes, it is
Suppose you estimate a regression model
not always crucial to report p-values for t statistics.
and obtain b^ 1 5 .56 and p-value 5 .086 for
But it does not hurt to report them. Further, when we
testing h0: b1 5 0 against h1: b1 2 0. What
discuss F testing in Section 4-5, we will see that it is
is the p-value for testing h0: b1 5 0 against
important to compute p-values, because critical valh1: b1 . 0?
ues for F tests are not so easily memorized.
4-2e A Reminder on the Language of Classical Hypothesis Testing
When H0 is not rejected, we prefer to use the language “we fail to reject H0 at the x% level,” rather
than “H0 is accepted at the x% level.” We can use Example 4.5 to illustrate why the former statement
is preferred. In this example, the estimated elasticity of price with respect to nox is 2.954, and the
t statistic for testing H0: bnox 5 21 is t 5 .393; therefore, we cannot reject H0. But there are many
other values for bnox (more than we can count) that cannot be rejected. For example, the t statistic
for H0: bnox 5 2.9 is 1 2.954 1 .9 2 /.117 5 2.462, and so this null is not rejected either. Clearly
bnox 5 21 and bnox 5 2.9 cannot both be true, so it makes no sense to say that we “accept” either of
these hypotheses. All we can say is that the data do not allow us to reject either of these hypotheses at
the 5% significance level.
4-2f Economic, or Practical, versus Statistical Significance
Because we have emphasized statistical significance throughout this section, now is a good time to
remember that we should pay attention to the magnitude of the coefficient estimates in addition to the
size of the t statistics. The statistical significance of a variable xj is determined entirely by the size of
tb^ j, whereas the economic significance or practical significance of a variable is related to the size
(and sign) of b^ j.
Recall that the t statistic for testing H0: bj 5 0 is defined by dividing the estimate by its standard error: tb^ j 5 b^ j/se 1 b^ j 2 . Thus, tb^ j can indicate statistical significance either because b^ j is “large”
or because se 1 b^ j 2 is “small.” It is important in practice to distinguish between these reasons for statistically significant t statistics. Too much focus on statistical significance can lead to the false conclusion that a variable is “important” for explaining y even though its estimated effect is modest.
58860_ch04_hr_117-162.indd 132
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
Example 4.6
133
Participation Rates in 401(k) Plans
In Example 3.3, we used the data on 401(k) plans to estimate a model describing participation rates in
terms of the firm’s match rate and the age of the plan. We now include a measure of firm size, the total
number of firm employees (totemp). The estimated equation is
prate 5 80.29 1 5.44 mrate 1 .269 age 2 .00013 totemp
1 0.78 2 1 0.52 2
1 .045 2
1 .00004 2
n 5 1,534, R2 5 .100.
The smallest t statistic in absolute value is that on the variable totemp: t 5 2.00013/.00004 5 23.25,
and this is statistically significant at very small significance levels. (The two-tailed p-value for this
t statistic is about .001.) Thus, all of the variables are statistically significant at rather small significance levels.
How big, in a practical sense, is the coefficient on totemp? Holding mrate and age fixed, if a firm
grows by 10,000 employees, the participation rate falls by 10,000 1 .00013 2 5 1.3 percentage points.
This is a huge increase in number of employees with only a modest effect on the participation rate.
Thus, although firm size does affect the participation rate, the effect is not practically very large.
The previous example shows that it is especially important to interpret the magnitude of the coefficient, in addition to looking at t statistics, when working with large samples. With large sample
sizes, parameters can be estimated very precisely: standard errors are often quite small relative to the
coefficient estimates, which usually results in statistical significance.
Some researchers insist on using smaller significance levels as the sample size increases, partly
as a way to offset the fact that standard errors are getting smaller. For example, if we feel comfortable
with a 5% level when n is a few hundred, we might use the 1% level when n is a few thousand. Using
a smaller significance level means that economic and statistical significance are more likely to coincide, but there are no guarantees: in the previous example, even if we use a significance level as small
as .1% (one-tenth of 1%), we would still conclude that totemp is statistically significant.
Many researchers are also willing to entertain larger significance levels in applications with
small sample sizes, reflecting the fact that it is harder to find significance with smaller sample sizes.
(Smaller sample sizes lead to less precise estimators, and the critical values are larger in magnitude,
two factors that make it harder to find statistical significance.) Unfortunately, one’s willingness to
consider higher significance levels can depend on one’s underlying agenda.
Example 4.7
Effect of Job Training on Firm Scrap Rates
The scrap rate for a manufacturing firm is the number of defective items—products that must be
­discarded—out of every 100 produced. Thus, for a given number of items produced, a decrease in the
scrap rate reflects higher worker productivity.
We can use the scrap rate to measure the effect of worker training on productivity. Using the data
in JTRAIN, but only for the year 1987 and for nonunionized firms, we obtain the following estimated
equation:
log 1 scrap 2 5 12.46 2 .029 hrsemp 2 .962 log 1 sales 2 1 .761 log 1 employ 2
1 5.69 2 1 .023 2
1 .453 2
1 .407 2
n 5 29, R2 5 .262.
The variable hrsemp is annual hours of training per employee, sales is annual firm sales (in dollars),
and employ is the number of firm employees. For 1987, the average scrap rate in the sample is about
4.6 and the average of hrsemp is about 8.9.
58860_ch04_hr_117-162.indd 133
10/19/18 8:32 PM
134
PART 1 Regression Analysis with Cross-Sectional Data
The main variable of interest is hrsemp. One more hour of training per employee lowers log(scrap)
by .029, which means the scrap rate is about 2.9% lower. Thus, if hrsemp increases by 5—each
employee is trained 5 more hours per year—the scrap rate is estimated to fall by 5 1 2.9 2 5 14.5%.
This seems like a reasonably large effect, but whether the additional training is worthwhile to the firm
depends on the cost of training and the benefits from a lower scrap rate. We do not have the numbers
needed to do a cost-benefit analysis, but the estimated effect seems nontrivial.
What about the statistical significance of the training variable? The t statistic on hrsemp is
2.029/.023 5 21.26, and now you probably recognize this as not being large enough in magnitude
to conclude that hrsemp is statistically significant at the 5% level. In fact, with 29 2 4 5 25 degrees
of freedom for the one-sided alternative, H1: bhrsemp , 0, the 5% critical value is about 21.71. Thus,
using a strict 5% level test, we must conclude that hrsemp is not statistically significant, even using a
one-sided alternative.
Because the sample size is pretty small, we might be more liberal with the significance level. The
10% critical value is 21.32, and so hrsemp is almost significant against the one-sided alternative at
the 10% level. The p-value is easily computed as P 1 T25 , 21.26 2 5 .110. This may be a low enough
p-value to conclude that the estimated effect of training is not just due to sampling error, but opinions
would legitimately differ on whether a one-sided p-value of .11 is sufficiently small.
Remember that large standard errors can also be a result of multicollinearity (high correlation
among some of the independent variables), even if the sample size seems fairly large. As we discussed in Section 3-4, there is not much we can do about this problem other than to collect more data
or change the scope of the analysis by dropping or combining certain independent variables. As in
the case of a small sample size, it can be hard to precisely estimate partial effects when some of the
explanatory variables are highly correlated. (Section 4-5 contains an example.)
We end this section with some guidelines for discussing the economic and statistical significance
of a variable in a multiple regression model:
1. Check for statistical significance. If the variable is statistically significant, discuss the magnitude
of the coefficient to get an idea of its practical or economic importance. This latter step can require
some care, depending on how the independent and dependent variables appear in the equation. (In
particular, what are the units of measurement? Do the variables appear in logarithmic form?)
2. If a variable is not statistically significant at the usual levels (10%, 5%, or 1%), you might still
ask if the variable has the expected effect on y and whether that effect is practically large. If it is
large, you should compute a p-value for the t statistic. For small sample sizes, you can sometimes
make a case for p-values as large as .20 (but there are no hard rules). With large p-values, that is,
small t statistics, we are treading on thin ice because the practically large estimates may be due to
sampling error: a different random sample could result in a very different estimate.
3. It is common to find variables with small t statistics that have the “wrong” sign. For practical purposes, these can be ignored: we conclude that the variables are statistically insignificant. A significant variable that has the unexpected sign and a practically large effect is much more troubling
and difficult to resolve. One must usually think more about the model and the nature of the data
to solve such problems. Often, a counterintuitive, significant estimate results from the omission
of a key variable or from one of the important problems we will discuss in Chapters 9 and 15.
4-3 Confidence Intervals
Under the CLM assumptions, we can easily construct a confidence interval (CI) for the population
parameter bj. Confidence intervals are also called interval estimates because they provide a range of
likely values for the population parameter, and not just a point estimate.
58860_ch04_hr_117-162.indd 134
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
135
Using the fact that 1 b^ j 2 bj 2 /se 1 b^ j 2 has a t distribution with n 2 k 2 1 degrees of freedom
[see (4.3)], simple manipulation leads to a CI for the unknown bj: a 95% confidence interval, given by
b^ j 6 c # se 1 b^ j 2 ,
[4.16]
where the constant c is the 97.5th percentile in a tn2k21 distribution. More precisely, the lower and
upper bounds of the confidence interval are given by
bj ; b^ j 2 c # se 1 b^ j 2
#
and
bj ; b^ j 1 c # se 1 b^ j 2 ,
respectively.
At this point, it is useful to review the meaning of a confidence interval. If random samples were
obtained over and over again, with bj and bj computed each time, then the (unknown) population value
#
bj would lie in the interval ( bj, bj) for 95% of the samples. Unfortunately, for the single sample that
#
we use to construct the CI, we do not know whether bj is actually contained in the interval. We hope
we have obtained a sample that is one of the 95% of all samples where the interval estimate contains
bj, but we have no guarantee.
Constructing a confidence interval is very simple when using current computing technology. Three
quantities are needed: b^ j , se 1 b^ j 2 , and c. The coefficient estimate and its standard error are reported by
any regression package. To obtain the value c, we must know the degrees of freedom, n 2 k 2 1, and
the level of confidence—95% in this case. Then, the value for c is obtained from the tn2k21 distribution.
As an example, for df 5 n 2 k 2 1 5 25, a 95% confidence interval for any bj is given by
3 b^ j 2 2.06 # se 1 b^ j 2 , b^ j 1 2.06 # se 1 b^ j 2 4 .
When n 2 k 2 1 . 120, the tn2k21 distribution is close enough to normal to use the 97.5th percentile in a standard normal distribution for constructing a 95% CI: b^ j 6 1.96 # se 1 b^ j 2 . In fact, when
n 2 k 2 1 . 50, the value of c is so close to 2 that we can use a simple rule of thumb for a 95% confidence interval: b^ j plus or minus two of its standard errors. For small degrees of freedom, the exact
percentiles should be obtained from the t tables.
It is easy to construct confidence intervals for any other level of confidence. For example, a 90% CI is obtained by choosing c to be the 95th percentile in the tn2k21 distribution. When
df 5 n 2 k 2 1 5 25, c 5 1.71, and so the 90% CI is b^ j 6 1.71 # se 1 b^ j 2 , which is necessarily narrower than the 95% CI. For a 99% CI, c is the 99.5th percentile in the t25 distribution. When df 5 25,
the 99% CI is roughly b^ j 6 2.79 # se 1 b^ j 2 , which is inevitably wider than the 95% CI.
Many modern regression packages save us from doing any calculations by reporting a 95% CI
along with each coefficient and its standard error. After a confidence interval is constructed, it is easy
to carry out two-tailed hypotheses tests. If the null hypothesis is H0: bj 5 aj, then H0 is rejected against
H1: bj 2 aj at (say) the 5% significance level if, and only if, aj is not in the 95% confidence interval.
Example 4.8
Model of R&D Expenditures
Economists studying industrial organization are interested in the relationship between firm size—
often measured by annual sales—and spending on research and development (R&D). Typically, a
constant elasticity model is used. One might also be interested in the ceteris paribus effect of the profit
margin—that is, profits as a percentage of sales—on R&D spending. Using the data in RDCHEM on
32 U.S. firms in the chemical industry, we estimate the following equation (with standard errors in
parentheses below the coefficients):
log 1 rd 2 5 24.38 1 1.084 log 1 sales 2 1 .0217 profmarg
1 .47 2 1 .060 2
1 .0128 2
2
n 5 32, R 5 .918.
58860_ch04_hr_117-162.indd 135
10/19/18 8:32 PM
136
PART 1 Regression Analysis with Cross-Sectional Data
The estimated elasticity of R&D spending with respect to firm sales is 1.084, so that, holding profit margin fixed, a 1% increase in sales is associated with a 1.084% increase in R&D spending. (Incidentally,
R&D and sales are both measured in millions of dollars, but their units of measurement have no effect on
the elasticity estimate.) We can construct a 95% confidence interval for the sales elasticity once we note
that the estimated model has n 2 k 2 1 5 32 2 2 2 1 5 29 degrees of freedom. From Table G.2, we
find the 97.5th percentile in a t29 distribution: c 5 2.045. Thus, the 95% confidence interval for blog1sales2
is 1.084 6 .060 1 2.045 2 , or about (.961,1.21). That zero is well outside this interval is hardly surprising: we expect R&D spending to increase with firm size. More interesting is that unity is included in
the 95% confidence interval for blog1sales2, which means that we cannot reject H0: blog1sales2 5 1 against
H1: blog1sales2 2 1 at the 5% significance level. In other words, the estimated R&D-sales elasticity is not
statistically different from 1 at the 5% level. (The estimate is not practically different from 1, either.)
The estimated coefficient on profmarg is also positive, and the 95% confidence interval for the
population parameter, bprofmarg, is .0217 6 .0128 1 2.045 2 , or about (2.0045, .0479). In this case, zero is
included in the 95% confidence interval, so we fail to reject H0: bprofmarg 5 0 against H1: bprofmarg 2 0
at the 5% level. Nevertheless, the t statistic is about 1.70, which gives a two-sided p-value of about .10,
and so we would conclude that profmarg is statistically significant at the 10% level against the twosided alternative, or at the 5% level against the one-sided alternative H1: bprofmarg . 0. Plus, the economic size of the profit margin coefficient is not trivial: holding sales fixed, a one percentage point
increase in profmarg is estimated to increase R&D spending by 100 1 .0217 2 < 2.2%. A complete
analysis of this example goes beyond simply stating whether a particular value, zero in this case, is or
is not in the 95% confidence interval.
You should remember that a confidence interval is only as good as the underlying assumptions used to construct it. If we have omitted important factors that are correlated with the explanatory variables, then the coefficient estimates are not reliable: OLS is biased. If heteroskedasticity
is ­present—for instance, in the previous example, if the variance of log(rd) depends on any of the
explanatory variables—then the standard error is not valid as an estimate of sd 1 b^ j 2 (as we discussed
in Section 3-4), and the confidence interval computed using these standard errors will not truly be a
95% CI. We have also used the normality assumption on the errors in obtaining these CIs, but, as we
will see in Chapter 5, this is not as important for applications involving hundreds of observations.
4-4 Testing Hypotheses about a Single Linear
Combination of the Parameters
The previous two sections have shown how to use classical hypothesis testing or confidence intervals
to test hypotheses about a single bj at a time. In applications, we must often test hypotheses involving
more than one of the population parameters. In this section, we show how to test a single hypothesis
involving more than one of the bj . Section 4-5 shows how to test multiple hypotheses.
To illustrate the general approach, we will consider a simple model to compare the returns to education at junior colleges and four-year colleges; for simplicity, we refer to the latter as “universities.”
[Kane and Rouse (1995) provide a detailed analysis of the returns to two- and four-year colleges.] The
population includes working people with a high school degree, and the model is
where
log 1 wage 2 5 b0 1 b1 jc 1 b2univ 1 b3exper 1 u,
[4.17]
jc 5 number of years attending a two-year college.
univ 5 number of years at a four-year college.
exper 5 months in the workforce.
58860_ch04_hr_117-162.indd 136
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
137
Note that any combination of junior college and four-year college is allowed, including jc 5 0 and
univ 5 0.
The hypothesis of interest is whether one year at a junior college is worth one year at a university:
this is stated as
H0: b 1 5 b 2.
[4.18]
Under H0, another year at a junior college and another year at a university lead to the same ceteris
paribus percentage increase in wage. For the most part, the alternative of interest is one-sided: a year
at a junior college is worth less than a year at a university. This is stated as
H1: b 1 , b 2.
[4.19]
The hypotheses in (4.18) and (4.19) concern two parameters, b1 and b2, a situation we have not
faced yet. We cannot simply use the individual t statistics for b^ 1 and b^ 2 to test H0. However, conceptually, there is no difficulty in constructing a t statistic for testing (4.18). To do so, we rewrite the
null and alternative as H0: b1 2 b2 5 0 and H1: b1 2 b2 , 0, respectively. The t statistic is based on
whether the estimated difference b^ 1 2 b^ 2 is sufficiently less than zero to warrant rejecting (4.18) in
favor of (4.19). To account for the sampling error in our estimators, we standardize this difference by
dividing by the standard error:
t5
b^ 1 2 b^ 2
.
se 1 b^ 1 2 b^ 2 2
[4.20]
Once we have the t statistic in (4.20), testing proceeds as before. We choose a significance level
for the test and, based on the df, obtain a critical value. Because the alternative is of the form in (4.19),
the rejection rule is of the form t , 2c, where c is a positive value chosen from the appropriate t distribution. Or we compute the t statistic and then compute the p-value (see Section 4-2).
The only thing that makes testing the equality of two different parameters more difficult than
testing about a single bj is obtaining the standard error in the denominator of (4.20). Obtaining the
numerator is trivial once we have performed the OLS regression. Using the data in TWOYEAR,
which comes from Kane and Rouse (1995), we estimate equation (4.17):
log 1 wage 2 5 1.472 1 .0667 jc 1 .0769 univ 1 .0049 exper
1 .021 2 1 .0068 2
1 .0023 2
1 .0002 2
[4.21]
2
n 5 6,763, R 5 .222.
It is clear from (4.21) that jc and univ have both economically and statistically significant effects on
wage. This is certainly of interest, but we are more concerned about testing whether the estimated difference in the coefficients is statistically significant. The difference is estimated as b^ 1 2 b^ 2 5 2.0102,
so the return to a year at a junior college is about one percentage point less than a year at a university.
Economically, this is not a trivial difference. The difference of 2.0102 is the numerator of the t statistic in (4.20).
Unfortunately, the regression results in equation (4.21) do not contain enough information to obtain
the standard error of b^ 1 2 b^ 2. It might be tempting to claim that se 1 b^ 1 2 b^ 2 2 5 se 1 b^ 1 2 2 se 1 b^ 2 2 , but
this is not true. In fact, if we reversed the roles of b^ 1 and b^ 2, we would wind up with a negative standard error of the difference using the difference in standard errors. Standard errors must always be
positive because they are estimates of standard deviations. Although the standard error of the difference b^ 1 2 b^ 2 certainly depends on se 1 b^ 1 2 and se 1 b^ 2 2 , it does so in a somewhat complicated way. To
find se 1 b^ 1 2 b^ 2 2 , we first obtain the variance of the difference. Using the results on variances in Math
Refresher B, we have
Var 1 b^ 1 2 b^ 2 2 5 Var 1 b^ 1 2 1 Var 1 b^ 2 2 2 2 Cov 1 b^ 1, b^ 2 2 .
58860_ch04_hr_117-162.indd 137
[4.22]
10/19/18 8:32 PM
138
PART 1 Regression Analysis with Cross-Sectional Data
Observe carefully how the two variances are added together, and twice the covariance is then subtracted. The standard deviation of b^ 1 2 b^ 2 is just the square root of (4.22), and, because 3 se 1 b^ 1 2 4 2 is
an unbiased estimator of Var 1 b^ 1 2 , and similarly for 3 se 1 b^ 2 2 4 2, we have
se 1 b^ 1 2 b^ 2 2 5 5 3 se 1 b^ 1 2 4 2 1 3 se 1 b^ 2 2 4 2 2 2s12 6 1/2,
[4.23]
H0: u 1 5 0 against H1: u 1 , 0.
[4.24]
where s12 denotes an estimate of Cov 1 b^ 1,b^ 2 2 . We have not displayed a formula for Cov 1 b^ 1,b^ 2 2 . Some
regression packages have features that allow one to obtain s12, in which case one can compute the
standard error in (4.23) and then the t statistic in (4.20). Advanced Treatment E shows how to use
matrix algebra to obtain s12.
Some of the more sophisticated econometrics programs include special commands that can be
used for testing hypotheses about linear combinations. Here, we cover an approach that is simple to
compute in virtually any statistical package. Rather than trying to compute se 1 b^ 1 2 b^ 2 2 from (4.23),
it is much easier to estimate a different model that directly delivers the standard error of interest.
Define a new parameter as the difference between b1 and b2: u 1 5 b1 2 b2. Then, we want to test
The t statistic in (4.20) in terms of u^ 1 is just t 5 u^ 1 /se 1 u^ 1 2 . The challenge is finding se 1 u^ 1 2 .
We can do this by rewriting the model so that u 1 appears directly on one of the independent
variables. Because u 1 5 b1 2 b2, we can also write b1 5 u 1 1 b2. Plugging this into (4.17) and rearranging gives the equation
log 1 wage 2 5 b0 1 1 u 1 1 b2 2 jc 1 b2univ 1 b3exper 1 u
5 b0 1 u 1 jc 1 b2 1 jc 1 univ 2 1 b3exper 1 u.
[4.25]
The key insight is that the parameter we are interested in testing hypotheses about, u 1, now multiplies
the variable jc. The intercept is still b0, and exper still shows up as being multiplied by b3. More
importantly, there is a new variable multiplying b2, namely jc 1 univ. Thus, if we want to directly
estimate u 1 and obtain the standard error of u^ 1, then we must construct the new variable jc 1 univ
and include it in the regression model in place of univ. In this example, the new variable has a natural
interpretation: it is total years of college, so define totcoll 5 jc 1 univ and write (4.25) as
log 1 wage 2 5 b0 1 u 1 jc 1 b2totcoll 1 b3exper 1 u.
[4.26]
The parameter b1 has disappeared from the model, while u 1 appears explicitly. This model is really
just a different way of writing the original model. The only reason we have defined this new model is
that, when we estimate it, the coefficient on jc is u^ 1, and, more importantly, se 1 u^ 1 2 is reported along
with the estimate. The t statistic that we want is the one reported by any regression package on the
variable jc (not the variable totcoll).
When we do this with the 6,763 observations used earlier, the result is
log 1 wage 2 5 1.472 2 .0102 jc 1 .0769 totcoll 1 .0049 exper
1 .021 2 1 .0069 2
1 .0023 2
1 .0002 2
2
n 5 6,763, R 5 .222.
[4.27]
The only number in this equation that we could not get from (4.21) is the standard error for the estimate 2.0102, which is .0069. The t statistic for testing (4.18) is 2.0102/.0069 5 21.48. Against
the one-sided alternative (4.19), the p-value is about .070, so there is some, but not strong, evidence
against (4.18).
The intercept and slope estimate on exper, along with their standard errors, are the same as in
(4.21). This fact must be true, and it provides one way of checking whether the transformed equation
has been properly estimated. The coefficient on the new variable, totcoll, is the same as the coefficient
on univ in (4.21), and the standard error is also the same. We know that this must happen by comparing (4.17) and (4.25).
58860_ch04_hr_117-162.indd 138
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
139
It is quite simple to compute a 95% confidence interval for u 1 5 b1 2 b2. Using the standard
normal approximation, the CI is obtained as usual: u^ 1 6 1.96 se 1 u^ 1 2 , which in this case leads to
2.0102 6 .0135.
The strategy of rewriting the model so that it contains the parameter of interest works in all cases
and is easy to implement. (See Computer Exercises C1 and C3 for other examples.)
4-5 Testing Multiple Linear Restrictions: The F Test
The t statistic associated with any OLS coefficient can be used to test whether the corresponding
unknown parameter in the population is equal to any given constant (which is usually, but not always,
zero). We have just shown how to test hypotheses about a single linear combination of the bj by rearranging the equation and running a regression using transformed variables. But so far, we have only
covered hypotheses involving a single restriction. Frequently, we wish to test multiple hypotheses
about the underlying parameters b0, b1, . . . , bk. We begin with the leading case of testing whether a
set of independent variables has no partial effect on a dependent variable.
4-5a Testing Exclusion Restrictions
We already know how to test whether a particular variable has no partial effect on the dependent variable: use the t statistic. Now, we want to test whether a group of variables has no effect on the dependent variable. More precisely, the null hypothesis is that a set of variables has no effect on y, once
another set of variables has been controlled.
As an illustration of why testing significance of a group of variables is useful, we consider the
following model that explains major league baseball players’ salaries:
log 1 salary 2 5 b0 1 b1years 1 b2gamesyr 1 b3bavg
1 b4hrunsyr 1 b5rbisyr 1 u,
[4.28]
H0: b3 5 0, b4 5 0, b5 5 0.
[4.29]
where salary is the 1993 total salary, years is years in the league, gamesyr is average games played
per year, bavg is career batting average (for example, bavg 5 250), hrunsyr is home runs per year,
and rbisyr is runs batted in per year. Suppose we want to test the null hypothesis that, once years in
the league and games per year have been controlled for, the statistics measuring performance—bavg,
hrunsyr, and rbisyr—have no effect on salary. Essentially, the null hypothesis states that productivity
as measured by baseball statistics has no effect on salary.
In terms of the parameters of the model, the null hypothesis is stated as
The null (4.29) constitutes three exclusion restrictions: if (4.29) is true, then bavg, hrunsyr, and
rbisyr have no effect on log(salary) after years and gamesyr have been controlled for and therefore
should be excluded from the model. This is an example of a set of multiple restrictions because we
are putting more than one restriction on the parameters in (4.28); we will see more general examples
of multiple restrictions later. A test of multiple restrictions is called a multiple hypotheses test or a
joint hypotheses test.
What should be the alternative to (4.29)? If what we have in mind is that “performance statistics
matter, even after controlling for years in the league and games per year,” then the appropriate alternative is simply
H1: H0 is not true.
[4.30]
The alternative (4.30) holds if at least one of b3, b4, or b5 is different from zero. (Any or all could be
different from zero.) The test we study here is constructed to detect any violation of H0. It is also valid
when the alternative is something like H1: b3 . 0, or b4 . 0, or b5 . 0, but it will not be the best
58860_ch04_hr_117-162.indd 139
10/19/18 8:32 PM
140
PART 1 Regression Analysis with Cross-Sectional Data
possible test under such alternatives. We do not have the space or statistical background necessary to
cover tests that have more power under multiple one-sided alternatives.
How should we proceed in testing (4.29) against (4.30)? It is tempting to test (4.29) by using the
t statistics on the variables bavg, hrunsyr, and rbisyr to determine whether each variable is individually significant. This option is not appropriate. A particular t statistic tests a hypothesis that puts no
restrictions on the other parameters. Besides, we would have three outcomes to contend with—one for
each t statistic. What would constitute rejection of (4.29) at, say, the 5% level? Should all three or only
one of the three t statistics be required to be significant at the 5% level? These are hard questions, and
fortunately we do not have to answer them. Furthermore, using separate t statistics to test a multiple
hypothesis like (4.29) can be very misleading. We need a way to test the exclusion restrictions jointly.
To illustrate these issues, we estimate equation (4.28) using the data in MLB1. This gives
log 1 salary 2 5 11.19 1 .0689 years 1 .0126 gamesyr
1 0.29 2 1 .0121 2
1 .0026 2
1 .00098 bavg 1 .0144 hrunsyr 1 .0108 rbisyr
1 .00110 2
1 .0161 2
1 .0072 2
n 5 353, SSR 5 183.186, R2 5 .6278,
[4.31]
where SSR is the sum of squared residuals. (We will use this later.) We have left several terms after the
decimal in SSR and R-squared to facilitate future comparisons. Equation (4.31) reveals that, whereas
years and gamesyr are statistically significant, none of the variables bavg, hrunsyr, and rbisyr has a
statistically significant t statistic against a two-sided alternative, at the 5% significance level. (The
t statistic on rbisyr is the closest to being significant; its two-sided p-value is .134.) Thus, based on the
three t statistics, it appears that we cannot reject H0.
This conclusion turns out to be wrong. To see this, we must derive a test of multiple restrictions
whose distribution is known and tabulated. The sum of squared residuals now turns out to provide a
very convenient basis for testing multiple hypotheses. We will also show how the R-squared can be
used in the special case of testing for exclusion restrictions.
Knowing the sum of squared residuals in (4.31) tells us nothing about the truth of the hypothesis in (4.29). However, the factor that will tell us something is how much the SSR increases when
we drop the variables bavg, hrunsyr, and rbisyr from the model. Remember that, because the OLS
estimates are chosen to minimize the sum of squared residuals, the SSR always increases when variables are dropped from the model; this is an algebraic fact. The question is whether this increase is
large enough, relative to the SSR in the model with all of the variables, to warrant rejecting the null
hypothesis.
The model without the three variables in question is simply
log 1 salary 2 5 b0 1 b1years 1 b2gamesyr 1 u.
[4.32]
In the context of hypothesis testing, equation (4.32) is the restricted model for testing (4.29); model
(4.28) is called the unrestricted model. The restricted model always has fewer parameters than the
unrestricted model.
When we estimate the restricted model using the data in MLB1, we obtain
log 1 salary 2 5 11.22 1 .0713 years 1 .0202 gamesyr
1 .11 2 1 .0125 2
1 .0013 2
n 5 353, SSR 5 198.311, R2 5 .5971.
[4.33]
As we surmised, the SSR from (4.33) is greater than the SSR from (4.31), and the R‑squared from the
restricted model is less than the R-squared from the unrestricted model. What we need to decide is
whether the increase in the SSR in going from the unrestricted model to the restricted model (183.186
to 198.311) is large enough to warrant rejection of (4.29). As with all testing, the answer depends on
the significance level of the test. But we cannot carry out the test at a chosen significance level until we
58860_ch04_hr_117-162.indd 140
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
141
have a statistic whose distribution is known, and can be tabulated, under H0. Thus, we need a way to
combine the information in the two SSRs to obtain a test statistic with a known distribution under H0.
Because it is no more difficult, we might as well derive the test for the general case. Write the
unrestricted model with k independent variables as
y 5 b0 1 b1x1 1 p 1 bkxk 1 u;
[4.34]
the number of parameters in the unrestricted model is k 1 1. (Remember to add one for the intercept.) Suppose that we have q exclusion restrictions to test: that is, the null hypothesis states that q
of the variables in (4.34) have zero coefficients. For notational simplicity, assume that it is the last q
variables in the list of independent variables: xk2q11, . . . , xk. (The order of the variables, of course, is
arbitrary and unimportant.) The null hypothesis is stated as
H0: bk2q11 5 0, . . . , bk 5 0,
[4.35]
which puts q exclusion restrictions on the model (4.34). The alternative to (4.35) is simply that it is
false; this means that at least one of the parameters listed in (4.35) is different from zero. When we
impose the restrictions under H0, we are left with the restricted model:
y 5 b0 1 b1x1 1 p 1 bk2q xk2q 1 u.
[4.36]
In this subsection, we assume that both the unrestricted and restricted models contain an intercept,
because that is the case most widely encountered in practice.
Now, for the test statistic itself. Earlier, we suggested that looking at the relative increase in the
SSR when moving from the unrestricted to the restricted model should be informative for testing the
hypothesis (4.35). The F statistic (or F ratio) is defined by
F;
G o i n g F u rt h e r 4 . 4
Consider relating individual performance
on a standardized test, score, to a variety
of other variables. School factors include
average class size, per-student expenditures, average teacher compensation, and
total school enrollment. Other variables
specific to the student are family income,
mother’s education, father’s education,
and number of siblings. The model is
score 5 b0 1 b1classize 1 b2expend
1 b3tchcomp 1 b4enroll
1 b5faminc 1 b6motheduc
1 b7fatheduc 1 b8siblings 1 u.
State the null hypothesis that studentspecific variables have no effect on standardized test performance once schoolrelated factors have been controlled for.
What are k and q for this example? Write
down the restricted version of the model.
1 SSRr 2 SSRur 2 /q
,
SSRur / 1 n 2 k 2 1 2
[4.37]
where SSRr is the sum of squared residuals from the
restricted model and SSRur is the sum of squared residuals from the unrestricted model.
You should immediately notice that, because SSRr
can be no smaller than SSRur, the F statistic is always
nonnegative (and almost always strictly positive). Thus,
if you compute a negative F statistic, then ­something is
wrong; the order of the SSRs in the numerator of F has
usually been reversed. Also, the SSR in the denominator of F is the SSR from the unrestricted model. The
easiest way to remember where the SSRs appear is to
think of F as measuring the relative increase in SSR
when moving from the unrestricted to the restricted
model.
The difference in SSRs in the numerator of
F is divided by q, which is the number of restrictions imposed in moving from the unrestricted to the
restricted model (q independent variables are dropped).
Therefore, we can write
q 5 numerator degrees of freedom 5 dfr 2 dfur, [4.38]
which also shows that q is the difference in degrees of freedom between the restricted and unrestricted models. (Recall that df 5 number of observations 2 number of estimated parameters. 2
58860_ch04_hr_117-162.indd 141
10/19/18 8:32 PM
142
PART 1 Regression Analysis with Cross-Sectional Data
Because the restricted model has fewer parameters—and each model is estimated using the same n
­observations—dfr is always greater than dfur.
The SSR in the denominator of F is divided by the degrees of freedom in the unrestricted model:
n 2 k 2 1 5 denominator degrees of freedom 5 dfur.
[4.39]
In fact, the denominator of F is just the unbiased estimator of s 5 Var 1 u 2 in the unrestricted model.
In a particular application, computing the F statistic is easier than wading through the somewhat
cumbersome notation used to describe the general case. We first obtain the degrees of freedom in the
unrestricted model, dfur. Then, we count how many variables are excluded in the restricted model; this
is q. The SSRs are reported with every OLS regression, and so forming the F statistic is simple.
In the major league baseball salary regression, n 5 353, and the full model (4.28) contains six
parameters. Thus, n 2 k 2 1 5 dfur 5 353 2 6 5 347. The restricted model (4.32) contains three
fewer independent variables than (4.28), and so q 5 3. Thus, we have all of the ingredients to compute the F statistic; we hold off doing so until we know what to do with it.
To use the F statistic, we must know its sampling distribution under the null in order to choose
critical values and rejection rules. It can be shown that, under H0 (and assuming the CLM assumptions hold), F is distributed as an F random variable with (q,n 2 k 2 1) degrees of freedom. We write
this as
2
F , Fq,n2k21.
The distribution of Fq,n2k21 is readily tabulated and available in statistical tables (see Table G.3) and,
even more importantly, in statistical software.
We will not derive the F distribution because the mathematics is very involved. Basically, it can
be shown that equation (4.37) is actually the ratio of two independent chi-square random variables,
divided by their respective degrees of freedom. The numerator chi-square random variable has q
degrees of freedom, and the chi-square in the denominator has n 2 k 2 1 degrees of freedom. This is
the definition of an F distributed random variable (see Math Refresher B).
It is pretty clear from the definition of F that we will reject H0 in favor of H1 when F is sufficiently “large.” How large depends on our chosen significance level. Suppose that we have decided on
a 5% level test. Let c be the 95th percentile in the Fq,n2k21 distribution. This critical value depends on
q (the numerator df ) and n 2 k 2 1 (the denominator df ). It is important to keep the numerator and
denominator degrees of freedom straight.
The 10%, 5%, and 1% critical values for the F distribution are given in Table G.3. The rejection
rule is simple. Once c has been obtained, we reject H0 in favor of H1 at the chosen significance level if
F . c.
[4.40]
With a 5% significance level, q 5 3, and n 2 k 2 1 5 60, the critical value is c 5 2.76. We would
reject H0 at the 5% level if the computed value of the F statistic exceeds 2.76. The 5% critical value
and rejection region are shown in Figure 4.7. For the same degrees of freedom, the 1% critical value
is 4.13.
In most applications, the numerator degrees of freedom (q) will be notably smaller than the
denominator degrees of freedom 1 n 2 k 2 1 2 . Applications where n 2 k 2 1 is small are unlikely
to be successful because the parameters in the unrestricted model will probably not be precisely estimated. When the denominator df reaches about 120, the F distribution is no longer sensitive to it.
(This is entirely analogous to the t distribution being well approximated by the standard normal distribution as the df gets large.) Thus, there is an entry in the table for the denominator df 5 `, and this
is what we use with large samples (because n 2 k 2 1 is then large). A similar statement holds for a
very large numerator df, but this rarely occurs in applications.
If H0 is rejected, then we say that xk2q11, . . . , xk are jointly statistically significant (or just
jointly significant) at the appropriate significance level. This test alone does not allow us to say
which of the variables has a partial effect on y; they may all affect y or maybe only one affects y.
58860_ch04_hr_117-162.indd 142
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
143
Figure 4.7 The 5% critical value and rejection region in an F3,60 distribution.
area = .95
area = .05
0
2.76
rejection
region
If the null is not rejected, then the variables are jointly insignificant, which often justifies dropping
them from the model.
For the major league baseball example with three numerator degrees of freedom and 347 denominator degrees of freedom, the 5% critical value is 2.60, and the 1% critical value is 3.78. We reject H0
at the 1% level if F is above 3.78; we reject at the 5% level if F is above 2.60.
We are now in a position to test the hypothesis that we began this section with: after controlling for years and gamesyr, the variables bavg, hrunsyr, and rbisyr have no effect on players’ salaries. In practice, it is easiest to first compute 1 SSRr 2 SSRur 2 /SSRur and to multiply the result by
1 n 2 k 2 1 2 /q; the reason the formula is stated as in (4.37) is that it makes it easier to keep the
numerator and denominator degrees of freedom straight. Using the SSRs in (4.31) and (4.33), we have
F5
1 198.311 2 183.186 2 347
#
< 9.55.
183.186
3
This number is well above the 1% critical value in the F distribution with 3 and 347 degrees of freedom, and so we soundly reject the hypothesis that bavg, hrunsyr, and rbisyr have no effect on salary.
The outcome of the joint test may seem surprising in light of the insignificant t statistics for
the three variables. What is happening is that the two variables hrunsyr and rbisyr are highly correlated, and this multicollinearity makes it difficult to uncover the partial effect of each variable; this
is reflected in the individual t statistics. The F statistic tests whether these variables (including bavg)
are jointly significant, and multicollinearity between hrunsyr and rbisyr is much less relevant for testing this hypothesis. In Computer Exercise C5, you are asked to reestimate the model while dropping
58860_ch04_hr_117-162.indd 143
10/19/18 8:32 PM
144
PART 1 Regression Analysis with Cross-Sectional Data
rbisyr, in which case hrunsyr becomes very significant. The same is true for rbisyr when hrunsyr is
dropped from the model.
The F statistic is often useful for testing exclusion of a group of variables when the variables
in the group are highly correlated. For example, suppose we want to test whether firm performance
affects the salaries of chief executive officers. There are many ways to measure firm performance, and
it probably would not be clear ahead of time which measures would be most important. Because measures of firm performance are likely to be highly correlated, hoping to find individually significant
measures might be asking too much due to multicollinearity. But an F test can be used to determine
whether, as a group, the firm performance variables affect salary.
4-5b Relationship between F and t Statistics
We have seen in this section how the F statistic can be used to test whether a group of variables
should be included in a model. What happens if we apply the F statistic to the case of testing significance of a single independent variable? This case is certainly not ruled out by the previous development. For example, we can take the null to be H0: bk 5 0 and q 5 1 (to test the single exclusion
restriction that xk can be excluded from the model). From Section 4-2, we know that the t statistic
on bk can be used to test this hypothesis. The question, then, is: do we have two separate ways of
testing hypotheses about a single coefficient? The answer is no. It can be shown that the F statistic for testing exclusion of a single variable is equal to the square of the corresponding t statistic.
Because t2n2k21 has an F1,n2k21 distribution, the two approaches lead to exactly the same outcome,
provided that the alternative is two-sided. The t statistic is more flexible for testing a single hypothesis because it can be directly used to test against one-sided alternatives. Because t statistics are
also easier to obtain than F statistics, there is really no reason to use an F statistic to test hypotheses
about a single parameter.
We have already seen in the salary regressions for major league baseball players that two (or
more) variables that each have insignificant t statistics can be jointly very significant. It is also possible that, in a group of several explanatory variables, one variable has a significant t statistic but the
group of variables is jointly insignificant at the usual significance levels. What should we make of
this kind of outcome? For concreteness, suppose that in a model with many explanatory variables we
cannot reject the null hypothesis that b1, b2, b3, b4, and b5 are all equal to zero at the 5% level, yet
the t statistic for b^ 1 is significant at the 5% level. Logically, we cannot have b1 2 0 but also have b1,
b2, b3, b4, and b5 all equal to zero! But as a matter of testing, it is possible that we can group a bunch
of insignificant variables with a significant variable and conclude that the entire set of variables is
jointly insignificant. (Such possible conflicts between a t test and a joint F test give another example
of why we should not “accept” null hypotheses; we should only fail to reject them.) The F statistic is
intended to detect whether a set of coefficients is different from zero, but it is never the best test for
determining whether a single coefficient is different from zero. The t test is best suited for testing a
single hypothesis. (In statistical terms, an F statistic for joint restrictions including b1 5 0 will have
less power for detecting b1 2 0 than the usual t statistic. See Section C-6 in Math Refresher C for a
discussion of the power of a test.)
Unfortunately, the fact that we can sometimes hide a statistically significant variable along with
some insignificant variables could lead to abuse if regression results are not carefully reported. For
example, suppose that, in a study of the determinants of loan-acceptance rates at the city level, x1
is the fraction of black households in the city. Suppose that the variables x2, x3, x4, and x5 are the
fractions of households headed by different age groups. In explaining loan rates, we would include
measures of income, wealth, credit ratings, and so on. Suppose that age of household head has no
effect on loan approval rates, once other variables are controlled for. Even if race has a marginally significant effect, it is possible that the race and age variables could be jointly insignificant.
Someone wanting to conclude that race is not a factor could simply report something like “Race
and age variables were added to the equation, but they were jointly insignificant at the 5% level.”
58860_ch04_hr_117-162.indd 144
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
145
Hopefully, peer review prevents these kinds of misleading conclusions, but you should be aware that
such outcomes are possible.
Often, when a variable is very statistically significant and it is tested jointly with another set of
variables, the set will be jointly significant. In such cases, there is no logical inconsistency in rejecting
both null hypotheses.
4-5c The R-Squared Form of the F Statistic
For testing exclusion restrictions, it is often more convenient to have a form of the F statistic that can
be computed using the R-squareds from the restricted and unrestricted models. One reason for this is
that the R-squared is always between zero and one, whereas the SSRs can be very large depending on
the unit of measurement of y, making the calculation based on the SSRs tedious. Using the fact that
SSRr 5 SST 1 1 2 R2r 2 and SSRur 5 SST 1 1 2 R2ur 2 , we can substitute into (4.37) to obtain
F5
1 R2ur 2 R2r 2 /q
1 R2ur 2 R2r 2 /q
5
1 1 2 R2ur 2 / 1 n 2 k 2 1 2
1 1 2 R2ur 2 /dfur
[4.41]
(note that the SST terms cancel everywhere). This is called the R-squared form of the F statistic.
[At this point, you should be cautioned that although equation (4.41) is very convenient for testing
exclusion restrictions, it cannot be applied for testing all linear restrictions. As we will see when we
discuss testing general linear restrictions, the sum of squared residuals form of the F statistic is sometimes needed.]
Because the R-squared is reported with almost all regressions (whereas the SSR is not), it is
easy to use the R-squareds from the unrestricted and restricted models to test for exclusion of some
variables. Particular attention should be paid to the order of the R-squareds in the numerator: the unrestricted R-squared comes first [contrast this with the SSRs in (4.37)]. Because R2ur . R2r , this shows
again that F will always be positive.
In using the R-squared form of the test for excluding a set of variables, it is important to not
square the R-squared before plugging it into formula (4.41); the squaring has already been done.
All regressions report R2, and these numbers are plugged directly into (4.41). For the baseball salary
example, we can use (4.41) to obtain the F statistic:
F5
1 .6278 2 .5971 2 347
#
< 9.54,
1 1 2 .6278 2
3
which is very close to what we obtained before. (The difference is due to rounding error.)
Example 4.9
Parents’ Education in a Birth Weight Equation
As another example of computing an F statistic, consider the following model to explain child birth
weight in terms of various factors:
bwght 5 b0 1 b1cigs 1 b2parity 1 b3faminc
1 b4motheduc 1 b5fatheduc 1 u,
[4.42]
where
bwght 5 birth weight, in pounds.
cigs 5 average number of cigarettes the mother smoked per day during pregnancy.
parity 5 the birth order of this child.
faminc 5 annual family income.
motheduc 5 years of schooling for the mother.
fatheduc 5 years of schooling for the father.
58860_ch04_hr_117-162.indd 145
10/19/18 8:32 PM
146
PART 1 Regression Analysis with Cross-Sectional Data
Let us test the null hypothesis that, after controlling for cigs, parity, and faminc, parents’ education
has no effect on birth weight. This is stated as H0: b4 5 0, b5 5 0, and so there are q 5 2 exclusion
restrictions to be tested. There are k 1 1 5 6 parameters in the unrestricted model (4.42); so the df
in the unrestricted model is n 2 6, where n is the sample size.
We will test this hypothesis using the data in BWGHT. This data set contains information on
1,388 births, but we must be careful in counting the observations used in testing the null hypothesis.
It turns out that information on at least one of the variables motheduc and fatheduc is missing for 197
births in the sample; these observations cannot be included when estimating the unrestricted model.
Thus, we really have n 5 1,191 observations, and so there are 1,191 2 6 5 1,185 df in the unrestricted model. We must be sure to use these same 1,191 observations when estimating the restricted
model (not the full 1,388 observations that are available). Generally, when estimating the restricted
model to compute an F test, we must use the same observations to estimate the unrestricted model;
otherwise, the test is not valid. When there are no missing data, this will not be an issue.
The numerator df is 2, and the denominator df is 1,185; from Table G.3, the 5% critical value
is c 5 3.0. Rather than report the complete results, for brevity, we present only the R-squareds.
The R-squared for the full model turns out to be R2ur 5 .0387. When motheduc and fatheduc are dropped from the regression, the R-squared falls to R2r 5 .0364. Thus, the F statistic is
F 5 3 1 .0387 2 .0364 2 / 1 1 2 .0387 2 4 1 1,185/2 2 5 1.42; because this is well below the 5% critical
value, we fail to reject H0. In other words, motheduc and fatheduc are jointly insignificant in the birth
weight equation. Most statistical packages have built-in commands for testing multiple hypotheses
after OLS estimation, and so one need not worry about making the mistake of running the two regressions on different data sets. Typically, the commands are applied after estimation of the unrestricted
model, which means the smaller subset of data is used whenever there are missing values on some
variables. Formulas for computing the F statistic using matrix algebra—see Advanced Treatment E—
do not require estimation of the restricted model.
4-5d Computing p-values for F Tests
For reporting the outcomes of F tests, p-values are especially useful. Because the F distribution depends
on the numerator and denominator df, it is difficult to get a feel for how strong or weak the evidence is
against the null hypothesis simply by looking at the
value of the F statistic and one or two critical values.
G o i n g F u rt h e r 4 . 5
In the F testing context, the p-value is defined as
The data in ATTEND were used to estimate
the two equations
atndrte 5 47.13 1 13.37 priGPA
1 2.87 2 1 1.09 2
n 5 680, R2 5 .183
and
atndrte 5 75.70 1 17.26 priGPA 2 1.72 ACT
1 3.88 2 1 1.08 2
1?2
n 5 680, R2 5 .291,
where, as always, standard errors are in
parentheses; the standard error for ACT is
missing in the second equation. What is the
t statistic for the coefficient on ACT ? (Hint:
First compute the F statistic for significance
of ACT.)
58860_ch04_hr_117-162.indd 146
p-value 5 P 1 ^ . F 2 ,
[4.43]
where, for emphasis, we let ^ denote an F random
variable with (q,n 2 k 2 1) degrees of freedom, and
F is the actual value of the test statistic. The p-value
still has the same interpretation as it did for t statistics: it is the probability of observing a value of F at
least as large as we did, given that the null hypothesis is true. A small p-value is evidence against H0.
For example, p-value 5 .016 means that the chance
of observing a value of F as large as we did when
the null hypothesis was true is only 1.6%; we usually
reject H0 in such cases. If the p-value 5 .314, then
the chance of observing a value of the F statistic as
large as we did under the null hypothesis is 31.4%.
Most would find this to be pretty weak evidence
against H0.
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
147
As with t testing, once the p-value has been computed, the F test can be carried out at any significance level. For example, if the p-value 5 .024, we reject H0 at the 5% significance level but not at
the 1% level.
The p-value for the F test in Example 4.9 is .238, and so the null hypothesis that bmotheduc and
bfatheduc are both zero is not rejected at even the 20% significance level.
Many econometrics packages have a built-in feature for testing multiple exclusion restrictions. These packages have several advantages over calculating the statistics by hand: we will less
likely make a mistake, p-values are computed automatically, and the problem of missing data, as in
Example 4.9, is handled without any additional work on our part.
4-5e The F Statistic for Overall Significance of a Regression
A special set of exclusion restrictions is routinely tested by most regression packages. These restrictions have the same interpretation, regardless of the model. In the model with k independent variables,
we can write the null hypothesis as
H0: x1, x2, . . . , xk do not help to explain y.
This null hypothesis is, in a way, very pessimistic. It states that none of the explanatory variables has
an effect on y. Stated in terms of the parameters, the null is that all slope parameters are zero:
H0: b1 5 b2 5 p 5 bk 5 0,
[4.44]
and the alternative is that at least one of the bj is different from zero. Another useful way of stating the
null is that H0: E 1 y 0 x1, x2, . . . , xk 2 5 E 1 y 2 , so that knowing the values of x1, x2, . . . , xk does not affect
the expected value of y.
There are k restrictions in (4.44), and when we impose them, we get the restricted model
y 5 b0 1 u;
[4.45]
all independent variables have been dropped from the equation. Now, the R-squared from estimating
(4.45) is zero; none of the variation in y is being explained because there are no explanatory variables.
Therefore, the F statistic for testing (4.44) can be written as
R2 /k
,
1 1 2 R2 2 / 1 n 2 k 2 1 2
[4.46]
where R2 is just the usual R-squared from the regression of y on x1, x2, . . . , xk.
Most regression packages report the F statistic in (4.46) automatically, which makes it tempting
to use this statistic to test general exclusion restrictions. You must avoid this temptation. The F statistic in (4.41) is used for general exclusion restrictions; it depends on the R-squareds from the restricted
and unrestricted models. The special form of (4.46) is valid only for testing joint exclusion of all independent variables. This is sometimes called determining the overall significance of the regression.
If we fail to reject (4.44), then there is no evidence that any of the independent variables help to
explain y. This usually means that we must look for other variables to explain y. For Example 4.9, the
F statistic for testing (4.44) is about 9.55 with k 5 5 and n 2 k 2 1 5 1,185 df. The p-value is zero to
four places after the decimal point, so that (4.44) is rejected very strongly. Thus, we conclude that the
variables in the bwght equation do explain some variation in bwght. The amount explained is not large:
only 3.87%. But the seemingly small R‑squared results in a highly significant F statistic. That is why we
must compute the F statistic to test for joint significance and not just look at the size of the R-squared.
Occasionally, the F statistic for the hypothesis that all independent variables are jointly insignificant is the focus of a study. Problem 10 asks you to use stock return data to test whether stock returns
over a four-year horizon are predictable based on information known only at the beginning of the
period. Under the efficient markets hypothesis, the returns should not be predictable; the null hypothesis is precisely (4.44).
58860_ch04_hr_117-162.indd 147
10/19/18 8:32 PM
148
PART 1 Regression Analysis with Cross-Sectional Data
4-5f Testing General Linear Restrictions
Testing exclusion restrictions is by far the most important application of F statistics. Sometimes, however, the restrictions implied by a theory are more complicated than just excluding some independent
variables. It is still straightforward to use the F statistic for testing.
As an example, consider the following equation:
where
log 1 price 2 5 b0 1 b1log 1 assess 2 1 b2log 1 lotsize 2
1 b3log 1 sqrft 2 1 b4bdrms 1 u,
[4.47]
price 5 house price.
assess 5 the assessed housing value 1 before the house was sold 2 .
lotsize 5 size of the lot, in square feet.
sqrft 5 square footage.
bdrms 5 number of bedrooms.
Now, suppose we would like to test whether the assessed housing price is a rational valuation. If
this is the case, then a 1% change in assess should be associated with a 1% change in price; that is,
b1 5 1. In addition, lotsize, sqrft, and bdrms should not help to explain log(price), once the assessed
value has been controlled for. Together, these hypotheses can be stated as
H0: b1 5 1, b2 5 0, b3 5 0, b4 5 0.
[4.48]
Four restrictions have to be tested; three are exclusion restrictions, but b1 5 1 is not. How can we test
this hypothesis using the F statistic?
As in the exclusion restriction case, we estimate the unrestricted model, (4.47) in this case, and
then impose the restrictions in (4.48) to obtain the restricted model. It is the second step that can be a
little tricky. But all we do is plug in the restrictions. If we write (4.47) as
y 5 b0 1 b1x1 1 b2x2 1 b3x3 1 b4x4 1 u,
[4.49]
then the restricted model is y 5 b0 1 x1 1 u. Now, to impose the restriction that the coefficient on x1
is unity, we must estimate the following model:
y 2 x1 5 b0 1 u.
[4.50]
This is just a model with an intercept 1 b0 2 but with a different dependent variable than in (4.49).
The procedure for computing the F statistic is the same: estimate (4.50), obtain the SSR 1 SSRr 2 ,
and use this with the unrestricted SSR from (4.49) in the F statistic (4.37). We are testing q 5 4 restrictions, and there are n 2 5 df in the unrestricted model. The F statistic is simply
3 1 SSRr 2 SSRur 2 /SSRur 4 3 1 n 2 5 2 /4 4 .
Before illustrating this test using a data set, we must emphasize one point: we cannot use the
R-squared form of the F statistic for this example because the dependent variable in (4.50) is different
from the one in (4.49). This means the total sum of squares from the two regressions will be different,
and (4.41) is no longer equivalent to (4.37). As a general rule, the SSR form of the F statistic should
be used if a different dependent variable is needed in running the restricted regression.
The estimated unrestricted model using the data in HPRICE1 is
log 1 price 2 5 .264 1 1.043 log 1 assess 2 1 .0074 log 1 lotsize 2
1 .570 2 1 .151 2
1 .0386 2
2 .1032 log 1 sqrft 2 1 .0338 bdrms
1 .1384 2
1 .0221 2
2
n 5 88, SSR 5 1.822, R 5 .773.
58860_ch04_hr_117-162.indd 148
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
149
If we use separate t statistics to test each hypothesis in (4.48), we fail to reject each one. But rationality of
the assessment is a joint hypothesis, so we should test the restrictions jointly. The SSR from the restricted
model turns out to be SSRr 5 1.880, and so the F statistic is 3 1 1.880 2 1.822 2 /1.822 4 1 83/4 2 5 .661.
The 5% critical value in an F distribution with (4,83) df is about 2.50, and so we fail to reject H0. There
is essentially no evidence against the hypothesis that the assessed values are rational.
4-6 Reporting Regression Results
We end this chapter by providing a few guidelines on how to report multiple regression results for
relatively complicated empirical projects. This should help you to read published works in the applied
social sciences, while also preparing you to write your own empirical papers. We will expand on this
topic in the remainder of the text by reporting results from various examples, but many of the key
points can be made now.
Naturally, the estimated OLS coefficients should always be reported. For the key variables in an
analysis, you should interpret the estimated coefficients (which often requires knowing the units of
measurement of the variables). For example, is an estimate an elasticity, or does it have some other
interpretation that needs explanation? The economic or practical importance of the estimates of the
key variables should be discussed.
The standard errors should always be included along with the estimated coefficients. Some
authors prefer to report the t statistics rather than the standard errors (and sometimes just the absolute value of the t statistics). Although nothing is really wrong with this, there is some preference for
reporting standard errors. First, it forces us to think carefully about the null hypothesis being tested;
the null is not always that the population parameter is zero. Second, having standard errors makes it
easier to compute confidence intervals.
The R-squared from the regression should always be included. We have seen that, in addition
to providing a goodness-of-fit measure, it makes calculation of F statistics for exclusion restrictions
simple. Reporting the sum of squared residuals and the standard error of the regression is sometimes
a good idea, but it is not crucial. The number of observations used in estimating any equation should
appear near the estimated equation.
If only a couple of models are being estimated, the results can be summarized in equation form,
as we have done up to this point. However, in many papers, several equations are estimated with many
different sets of independent variables. We may estimate the same equation for different groups of
people, or even have equations explaining different dependent variables. In such cases, it is better to
summarize the results in one or more tables. The dependent variable should be indicated clearly in the
table, and the independent variables should be listed in the first column. Standard errors (or t statistics) can be put in parentheses below the estimates.
Example 4.10
Salary-Pension Tradeoff for Teachers
Let totcomp denote average total annual compensation for a teacher, including salary and all fringe
benefits (pension, health insurance, and so on). Extending the standard wage equation, total compensation should be a function of productivity and perhaps other characteristics. As is standard, we use
logarithmic form:
log 1 totcomp 2 5 f 1 productivity characteristics, other factors 2 ,
where f 1 # 2 is some function (unspecified for now). Write
totcomp 5 salary 1 benefits 5 salary a1 1
58860_ch04_hr_117-162.indd 149
benefits
b.
salary
10/19/18 8:32 PM
150
PART 1 Regression Analysis with Cross-Sectional Data
This equation shows that total compensation is the product of two terms: salary and 1 1 b/s,
where b/s is shorthand for the “benefits to salary ratio.” Taking the log of this equation gives
log 1 totcomp 2 5 log 1 salary 2 1 log 1 1 1 b/s 2 . Now, for “small” b/s, log 1 1 1 b/s 2 < b/s; we will
use this approximation. This leads to the econometric model
log 1 salary 2 5 b0 1 b1 1 b/s 2 1 other factors.
Testing the salary-benefits tradeoff then is the same as a test of H0: b1 5 21 against H1: b1 2 21.
We use the data in MEAP93 to test this hypothesis. These data are averaged at the school level,
and we do not observe very many other factors that could affect total compensation. We will include
controls for size of the school (enroll ), staff per thousand students (staff ), and measures such as the
school dropout and graduation rates. The average b/s in the sample is about .205, and the largest value
is .450.
The estimated equations are given in Table 4.1, where standard errors are given in parentheses
below the coefficient estimates. The key variable is b/s, the benefits-salary ratio.
From the first column in Table 4.1, we see that, without controlling for any other factors, the
OLS coefficient for b/s is 2.825. The t statistic for testing the null hypothesis H0: b1 5 21 is
t 5 1 2.825 1 1 2 /.200 5 .875, and so the simple
regression fails to reject H0. After adding controls for
school size and staff size (which roughly captures the
G o i n g F u rt h e r 4 . 6
number of students taught by each teacher), the estiHow does adding droprate and gradrate
mate of the b/s coefficient becomes 2.605. Now, the
affect the estimate of the salary-benefits
test
of b1 5 21 gives a t statistic of about 2.39; thus,
tradeoff? Are these variables jointly signifiH0 is rejected at the 5% level against a two‑sided
cant at the 5% level? What about the 10%
alternative. The variables log(enroll) and log(staff )
level?
are very statistically significant.
Table 4.1 Testing the Salary-benefits Tradeoff
Dependent Variable: log(salary)
Independent Variables
(1)
(2)
b/s
–.825
(.200)
–.605
(.165)
–.589
(.165)
log(enroll)
——
.0874
(.0073)
.0881
(.0073)
log(staff )
——
–.222
(.050)
–.218
(.050)
droprate
——
——
–.00028
(.00161)
gradrate
——
——
.00097
.00066)
intercept
10.523
(0.042)
10.884
(0.252)
10.738
(0.258)
408
.040
408
.353
408
.361
Observations
R-squared
58860_ch04_hr_117-162.indd 150
(3)
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
151
4-7 Revisiting Causal Effects and Policy Analysis
In Section 3-7e we showed how multiple regression can be used to obtain unbiased estimators of
causal, or treatment, effects in the context of policy interventions, provided we have controls sufficient to ensure that participation assignment is unconfounded. In particular, with a constant treatment
effect, t, we derived
E(y 0 w, x) 5 a 1 tw 1 xg 5 a 1 tw 1 g1x1 1 ? ? ? 1 gk xk ,
where y is the outcome or response, w is the binary policy (treatment) variable, and the xj are the
controls that account for nonrandom assignment. We know that the OLS estimator of t is unbiased because MLR.1 and MLR.4 hold (and we have random sampling from the population). If we
add MLR.5 and MLR.6, we can perform exact inference on t. For example, the null hypothesis
of no policy effect is H0 : t 5 0, and we can test this hypothesis—against a one-sided or two-sided
­alternative—using a standard t statistic. Regardless of the magnitude of the estimate t^ , most researchers and administrators will not be convinced that an intervention or policy is effective unless t^ is
statistically different from zero (and with the expected sign) at a sufficiently small significance level.
As in any context, it is important to discuss the sign and magnitude of t^ in addition to its statistical
significance. Probably of more interest is to obtain a 95% confidence interval for t, which gives us a
plausible range of ­values for the population treatment effect.
We can also test hypotheses about the gj, but, in a policy environment, we are rarely concerned
about the statistical significance of the xj except perhaps as a logical check on the regression results. For
example, we should expect past labor market earnings to positively predict current labor market earnings.
We now revisit Example 3.7, which contains the estimated effect of a job training program using
JTRAIN98.
EXAMPLE 4.11
Evaluating a Job Training Program
We reproduce the simple and multiple regression estimates and now put the standard errors below the
coefficients. Recall that the outcome variable, earn98, is measured in thousands of dollars:
earn98 5 10.61 2 2.05 train
(0.28) (0.48)
n 5 1,130, R2 5 0.016
earn98 5 4.67 1 2.41 train 1 .373 earn96 1 .363 educ 2 .181 age 1 2.48 married
(1.15) (0.44)
(.019)
(.064)
(.019)
(0.43)
2
n 5 1,130, R 5 0. 405
[4.51]
[4.52]
As discussed in Example 3.7, the change in the sign of coefficient on train is striking when moving from simple to multiple regression. Moreover, the t statistic in (4.51) is 22.05/0.48 < 24.27,
which gives a very statistically significant and practically large negative effect of the program. By
contrast, the t statistic in (4.52) is about 5.47, which shows a strongly statistically significant and
­positive effect. It is pretty clear that we prefer the multiple regression results for evaluating the job
training program. Of course, it could be that we have omitted some important controls in (4.52), but at
a minimum we know that we can account for some important differences across workers.
Perhaps now is a good time to revisit the the multicollinearity issue, which we raised in
Section 3-4a. Recall that collinearity arises only in the context of multiple regression, and so our discussion is relevant only for equation (4.52). In this equation, it could be that two or more of the control
variables in (4.52) are highly correlated; or maybe not. The point is that we do not care. The reason we
58860_ch04_hr_117-162.indd 151
10/19/18 8:32 PM
152
PART 1 Regression Analysis with Cross-Sectional Data
include earn96, educ, age, and married is to control for differences in men that at least partly determine
participation in the job training program, hopefully leading to an unbiased estimator of the treatment
effect. We are not worried about how well we estimate the coefficients on the control variables, and
including highly correlated variables among the xj has nothing to do with obtaining a reliable estimate
of t. Computer Exercise C14 asks you to add a binary variable for whether the man was unemployed in
1996, which is strongly related to earnings in 1996: unem96 5 0 means earn96 5 0. And yet, correlation between unem96 and earn96 is of essentially no concern. If we could observe earnings in 1995,
earn95, we would likely include it, too, even though it is likely to be highly correlated with earn96.
Summary
In this chapter, we have covered the very important topic of statistical inference, which allows us to infer
something about the population model from a random sample. We summarize the main points:
1. Under the classical linear model assumptions MLR.1 through MLR.6, the OLS estimators are
­normally distributed.
2. Under the CLM assumptions, the t statistics have t distributions under the null hypothesis.
3. We use t statistics to test hypotheses about a single parameter against one- or two-sided alternatives,
using one- or two-tailed tests, respectively. The most common null hypothesis is H0: bj 5 0, but we
sometimes want to test other values of bj under H0.
4. In classical hypothesis testing, we first choose a significance level, which, along with the df and alternative hypothesis, determines the critical value against which we compare the t statistic. It is more
informative to compute the p-value for a t test—the smallest significance level for which the null
hypothesis is rejected—so that the hypothesis can be tested at any significance level.
5. Under the CLM assumptions, confidence intervals can be constructed for each bj. These CIs can be
used to test any null hypothesis concerning bj against a two-sided alternative.
6. Single hypothesis tests concerning more than one bj can always be tested by rewriting the model to
contain the parameter of interest. Then, a standard t statistic can be used.
7. The F statistic is used to test multiple exclusion restrictions, and there are two equivalent forms of the
test. One is based on the SSRs from the restricted and unrestricted models. A more convenient form is
based on the R-squareds from the two models.
8. When computing an F statistic, the numerator df is the number of restrictions being tested, while the
denominator df is the degrees of freedom in the unrestricted model.
9. The alternative for F testing is two-sided. In the classical approach, we specify a significance level
which, along with the numerator df and the denominator df, determines the critical value. The null
hypothesis is rejected when the statistic, F, exceeds the critical value, c. Alternatively, we can compute
a p-value to summarize the evidence against H0.
10. General multiple linear restrictions can be tested using the sum of squared residuals form of the
F statistic.
11. The F statistic for the overall significance of a regression tests the null hypothesis that all slope parameters are zero, with the intercept unrestricted. Under H0, the explanatory variables have no effect on
the expected value of y.
12. When data are missing on one or more explanatory variables, one must be careful when computing
F statistics “by hand,” that is, using either the sum of squared residuals or R-squareds from the two
regressions. Whenever possible it is best to leave the calculations to statistical packages that have
built-in commands, which work with or without missing data.
13. Statistical inference is important for program evaluation and policy analysis. Rarely is it enough to
report only the economic (or practical) significance of our estimates. We, and others, must be convinced that moderate to large estimates of treatment effects are not due purely to sampling variation.
Obtaining a p-value for the null hypothesis that the effect is zero, or, even better, obtaining a 95%
confidence interval, allows us to determine statistic significance in addition to economic significance.
58860_ch04_hr_117-162.indd 152
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
153
The Classical Linear Model Assumptions
Now is a good time to review the full set of classical linear model (CLM) assumptions for cross-sectional
regression. Following each assumption is a comment about its role in multiple regression analysis.
Assumption MLR.1 (Linear in Parameters)
The model in the population can be written as
y 5 b0 1 b1x1 1 b2 x2 1 p 1 bk xk 1 u,
where b0, b1, . . . , bk are the unknown parameters (constants) of interest and u is an unobserved random
­error or disturbance term.
Assumption MLR.1 describes the population relationship we hope to estimate, and explicitly sets out
the bj—the ceteris paribus population effects of the xj on y—as the parameters of interest.
Assumption MLR.2 (Random Sampling)
We have a random sample of n observations, 5 1 xi1, xi2, . . . , xik, yi 2 : i 5 1, . . . , n 6 , following the population model in Assumption MLR.1.
This random sampling assumption means that we have data that can be used to estimate the bj, and
that the data have been chosen to be representative of the population described in Assumption MLR.1.
Assumption MLR.3 (No Perfect Collinearity)
In the sample (and therefore in the population), none of the independent variables is constant, and there are
no exact linear relationships among the independent variables.
Once we have a sample of data, we need to know that we can use the data to compute the OLS estimates, the b^ j. This is the role of Assumption MLR.3: if we have sample variation in each independent variable and no exact linear relationships among the independent variables, we can compute the b^ j.
Assumption MLR.4 (Zero Conditional Mean)
The error u has an expected value of zero given any values of the explanatory variables. In other words,
E 1 u 0 x1, x2, . . . , xk 2 5 0.
As we discussed in the text, assuming that the unobserved factors are, on average, unrelated to the
explanatory variables is key to deriving the first statistical property of each OLS estimator: its unbiasedness
for the corresponding population parameter. Of course, all of the previous assumptions are used to show
unbiasedness.
Assumption MLR.5 (Homoskedasticity)
The error u has the same variance given any values of the explanatory variables. In other words,
Var 1 u 0 x1, x2, . . . , xk 2 5 s2.
Compared with Assumption MLR.4, the homoskedasticity assumption is of secondary importance; in
particular, Assumption MLR.5 has no bearing on the unbiasedness of the b^ j. Still, homoskedasticity has
two important implications: (1) We can derive formulas for the sampling variances whose components are
easy to characterize; (2) We can conclude, under the Gauss-Markov assumptions MLR.1 through MLR.5,
that the OLS estimators have smallest variance among all linear, unbiased estimators.
Assumption MLR.6 (Normality)
The population error u is independent of the explanatory variables x1, x2, . . . , xk and is normally distributed
with zero mean and variance s2: u , Normal 1 0, s2 2 .
In this chapter, we added Assumption MLR.6 to obtain the exact sampling distributions of t statistics and F statistics, so that we can carry out exact hypotheses tests. In the next chapter, we will see that
MLR.6 can be dropped if we have a reasonably large sample size. Assumption MLR.6 does imply a stronger ­efficiency property of OLS: the OLS estimators have smallest variance among all unbiased estimators;
the comparison group is no longer restricted to estimators linear in the 5 yi: i 5 1, 2, . . . , n 6 .
58860_ch04_hr_117-162.indd 153
10/19/18 8:32 PM
154
PART 1 Regression Analysis with Cross-Sectional Data
Key Terms
Alternative Hypothesis
Classical Linear Model
Classical Linear Model (CLM)
Assumptions
Confidence Interval (CI)
Critical Value
Denominator Degrees of Freedom
Economic Significance
Exclusion Restrictions
F Statistic
Joint Hypotheses Test
Jointly Insignificant
Jointly Statistically Significant
Minimum Variance Unbiased
Estimators
Multiple Hypotheses Test
Multiple Restrictions
Normality Assumption
Null Hypothesis
Numerator Degrees of Freedom
One-Sided Alternative
One-Tailed Test
Overall Significance of the Regression
p-value
Practical Significance
R-squared Form of the F Statistic
Rejection Rule
Restricted Model
Significance Level
Statistically Insignificant
Statistically Significant
t Ratio
t Statistic
Two-Sided Alternative
Two-Tailed Test
Unrestricted Model
Problems
1 Which of the following can cause the usual OLS t statistics to be invalid (that is, not to have t distributions under H0)?
(i) Heteroskedasticity.
(ii) A sample correlation coefficient of .95 between two independent variables that are in the model.
(iii) Omitting an important explanatory variable.
2 Consider an equation to explain salaries of CEOs in terms of annual firm sales, return on equity (roe, in
percentage form), and return on the firm’s stock (ros, in percentage form):
(i)
(ii)
log 1 salary 2 5 b0 1 b1log 1 sales 2 1 b2roe 1 b3ros 1 u.
In terms of the model parameters, state the null hypothesis that, after controlling for sales and
roe, ros has no effect on CEO salary. State the alternative that better stock market performance
increases a CEO’s salary.
Using the data in CEOSAL1, the following equation was obtained by OLS:
log 1 salary 2 5 4.32 1 .280 log 1 sales 2 1 .0174 roe 1 .00024 ros
1 .32 2 1 .035 2
2
n 5 209, R 5 .283.
1 .0041 2
1 .00054 2
By what percentage is salary predicted to increase if ros increases by 50 points? Does ros have
a practically large effect on salary?
(iii) Test the null hypothesis that ros has no effect on salary against the alternative that ros has a
positive effect. Carry out the test at the 10% significance level.
(iv) Would you include ros in a final model explaining CEO compensation in terms of firm performance? Explain.
3 The variable rdintens is expenditures on research and development (R&D) as a percentage of sales.
Sales are measured in millions of dollars. The variable profmarg is profits as a percentage of sales.
Using the data in RDCHEM for 32 firms in the chemical industry, the following equation
is estimated:
rdintens 5 .472 1 .321 log 1 sales 2 1 .050 profmarg
1 1.369 2 1 .216 2
2
n 5 32, R 5 .099.
(i)
58860_ch04_hr_117-162.indd 154
1 .046 2
Interpret the coefficient on log(sales). In particular, if sales increases by 10%, what is the estimated percentage point change in rdintens? Is this an economically large effect?
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
155
(ii)
Test the hypothesis that R&D intensity does not change with sales against the alternative that it
does increase with sales. Do the test at the 5% and 10% levels.
(iii) Interpret the coefficient on profmarg. Is it economically large?
(iv) Does profmarg have a statistically significant effect on rdintens ?
4 Are rent rates influenced by the student population in a college town? Let rent be the average monthly
rent paid on rental units in a college town in the United States. Let pop denote the total city population,
avginc the average city income, and pctstu the student population as a percentage of the total population. One model to test for a relationship is
(i)
log 1 rent 2 5 b0 1 b1log 1 pop 2 1 b2log 1 avginc 2 1 b3pctstu 1 u.
State the null hypothesis that size of the student body relative to the population has no ceteris
paribus effect on monthly rents. State the alternative that there is an effect.
(ii) What signs do you expect for b1 and b2?
(iii) The equation estimated using 1990 data from RENTAL for 64 college towns is
log 1 rent 2 5 .043 1 .066 log 1 pop 2 1 .507 log 1 avginc 2 1 .0056 pctstu
1 .844 2 1 .039 2
2
n 5 64, R 5 .458.
(iv)
1 .081 2
1 .0017 2
What is wrong with the statement: “A 10% increase in population is associated with about a
6.6% increase in rent”?
Test the hypothesis stated in part (i) at the 1% level.
5 Consider the estimated equation from Example 4.3, which can be used to study the effects of skipping
class on college GPA:
colGPA 5 1.39 1 .412 hsGPA 1 .015 ACT 2 .083 skipped
1 .33 2 1 .094 2
2
n 5 141, R 5 .234.
1 .011 2
1 .026 2
(i) Using the standard normal approximation, find the 95% confidence interval for bhsGPA.
(ii) Can you reject the hypothesis H0: bhsGPA 5 .4 against the two-sided alternative at the 5% level?
(iii) Can you reject the hypothesis H0: bhsGPA 5 1 against the two-sided alternative at the 5% level?
6 In Section 4-5, we used as an example testing the rationality of assessments of housing prices. There,
we used a log-log model in price and assess [see equation (4.47)]. Here, we use a level-level formulation.
(i) In the simple regression model
price 5 b0 1 b1assess 1 u,
the assessment is rational if b1 5 1 and b0 5 0. The estimated equation is
price 5 214.47 1 .976 assess
1 16.27 2 1 .049 2
n 5 88, SSR 5 165,644.51, R2 5 .820.
First, test the hypothesis that H0: b0 5 0 against the two-sided alternative. Then, test H0: b1 5 1
against the two-sided alternative. What do you conclude?
(ii) To test the joint hypothesis that b0 5 0 and b1 5 1, we need the SSR in the restricted model.
n
This amounts to computing g i51 1 pricei 2 assessi 2 2, where n 5 88, because the residuals in
the restricted model are just pricei 2 assessi. (No estimation is needed for the restricted model
because both parameters are specified under H0.) This turns out to yield SSR 5 209,448.99.
Carry out the F test for the joint hypothesis.
(iii) Now, test H0: b2 5 0, b3 5 0, and b4 5 0 in the model
price 5 b0 1 b1assess 1 b2lotsize 1 b3sqrft 1 b4bdrms 1 u.
The R-squared from estimating this model using the same 88 houses is .829.
58860_ch04_hr_117-162.indd 155
10/19/18 8:32 PM
156
PART 1 Regression Analysis with Cross-Sectional Data
(iv)
If the variance of price changes with assess, lotsize, sqrft, or bdrms, what can you say about the
F test from part (iii)?
7 In Example 4.7, we used data on nonunionized manufacturing firms to estimate the relationship between
the scrap rate and other firm characteristics. We now look at this example more closely and use all available firms.
(i) The population model estimated in Example 4.7 can be written as
log 1 scrap 2 5 b0 1 b1hrsemp 1 b2log 1 sales 2 1 b3log 1 employ 2 1 u.
Using the 43 observations available for 1987, the estimated equation is
log 1 scrap 2 5 11.74 2 .042 hrsemp 2 .951 log 1 sales 2 1 .992 log 1 employ 2
1 4.57 2 1 .019 2
2
n 5 43, R 5 .310.
(ii)
1 .370 2
1 .360 2
Compare this equation to that estimated using only the 29 nonunionized firms in the sample.
Show that the population model can also be written as
log 1 scrap 2 5 b0 1 b1hrsemp 1 b2log 1 sales/employ 2 1 u 3log 1 employ 2 1 u,
where u 3 5 b2 1 b3. [Hint: Recall that log 1 x2/x3 2 5 log 1 x2 2 2 log 1 x3 2 .] Interpret the hypothesis H0: u 3 5 0.
(iii) When the equation from part (ii) is estimated, we obtain
log 1 scrap 2 5 11.74 2 .042 hrsemp 2 .951 log 1 sales/employ 2 1 .041 log 1 employ 2
1 4.57 2 1 .019 2
2
n 5 43, R 5 .310.
(iv)
1 .370 2
1 .205 2
Controlling for worker training and for the sales-to-employee ratio, do bigger firms have larger
statistically significant scrap rates?
Test the hypothesis that a 1% increase in sales/employ is associated with a 1% drop in the scrap rate.
8 Consider the multiple regression model with three independent variables, under the classical linear
model assumptions MLR.1 through MLR.6:
y 5 b0 1 b1x1 1 b2x2 1 b3x3 1 u.
You would like to test the null hypothesis H0: b1 2 3b2 5 1.
(i) Let b^ 1 and b^ 2 denote the OLS estimators of b1 and b2. Find Var 1 b^ 1 2 3b^ 2 2 in terms of the variances of b^ 1 and b^ 2 and the covariance between them. What is the standard error of b^ 1 2 3b^ 2?
(ii) Write the t statistic for testing H0: b1 2 3b2 5 1.
(iii) Define u 1 5 b1 2 3b2 and u^ 1 5 b^ 1 2 3b^ 2. Write a regression equation involving b0, u 1, b2,
and b3 that allows you to directly obtain u^ 1 and its standard error.
9 In Problem 3 in Chapter 3, we estimated the equation
sleep 5 3,638.25 2 .148 totwrk 2 11.13 educ 1 2.20 age
1 112.28 2 1 .017 2
2
n 5 706, R 5 .113,
1 5.88 2
1 1.45 2
where we now report standard errors along with the estimates.
(i) Is either educ or age individually significant at the 5% level against a two-sided alternative?
Show your work.
(ii) Dropping educ and age from the equation gives
58860_ch04_hr_117-162.indd 156
10/19/18 8:32 PM
CHAPTER 4 Multiple Regression Analysis: Inference
157
sleep 5 3,586.38 2 .151 totwrk
1 38.91 2 1 .017 2
n 5 706, R2 5 .103.
Are educ and age jointly significant in the original equation at the 5% level? Justify your
answer.
(iii) Does including educ and age in the model greatly affect the estimated tradeoff between sleeping
and working?
(iv) Suppose that the sleep equation contains heteroskedasticity. What does this mean about the tests
computed in parts (i) and (ii)?
10 Regression analysis can be used to test whether the market efficiently uses information in valuing
stocks. For concreteness, let return be the total return from holding a firm’s stock over the four-year
period from the end of 1990 to the end of 1994. The efficient markets hypothesis says that these returns
should not be systematically related to information known in 1990. If firm characteristics known at the
beginning of the period help to predict stock returns, then we could use this information in choosing
stocks.
For 1990, let dkr be a firm’s debt to capital ratio, let eps denote the earnings per share, let netinc
denote net income, and let salary denote total compensation for the CEO.
(i) Using the data in RETURN, the following equation was estimated:
return 5 214.37 1 .321 dkr 1 .043 eps 2 .0051 nentinc 1 .0035 salary
1 6.89 2 1 .201 2
n 5 142, R2 5 .0395.
(ii)
1 .078 2
1 .0047 2
1 .0022 2
Test whether the explanatory variables are jointly significant at the 5% level. Is any explanatory
variable individually significant?
Now, reestimate the model using the log form for netinc and salary:
return 5 236.30 1 .327 dkr 1 .069 eps 2 4.74 log 1 netinc 2 1 7.24 log 1 salary 2
1 39.37 2 1 .203 2
2
n 5 142, R 5 .0330.
1 .080 2
1 3.39 2
1 6.31 2
Do any of your conclusions from part (i) change?
(iii) In this sample, some firms have zero debt and others have negative earnings. Should we try to
use log(dkr) or log(eps) in the model to see if these improve the fit? Explain.
(iv) Overall, is the evidence for predictability of stock returns strong or weak?
11 The following table was created using the data in CEOSAL2, where standard errors are in parentheses
below the coefficients:
Dependent Variable: log(salary)
Independent Variables
58860_ch04_hr_117-162.indd 157
(1)
(2)
(3)
log(sales)
.224
(.027)
.158
(.040)
.188
(.040)
log(mktval)
——
.112
(.050)
.100
(.049)
profmarg
——
–.0023
(.0022)
–.0022
(.0021)
10/19/18 8:32 PM
158
PART 1 Regression Analysis with Cross-Sectional Data
Dependent Variable: log(salary)
Independent Variables
(1)
(2)
(3)
ceoten
——
——
.0171
(.0055)
comten
——
——
–.0092
(.0033)
intercept
4.94
(0.20)
4.62
(0.25)
4.57
(0.25)
Observations
R-squared
177
.281
177
.304
177
.353
The variable mktval is market value of the firm, profmarg is profit as a percentage of sales, ceoten is
years as CEO with the current company, and comten is total years with the company.
(i) Comment on the effect of profmarg on CEO salary.
(ii) Does market value have a significant effect? Explain.
(iii) Interpret the coefficients on ceoten and comten. Are these explanatory variables statistically
significant?
(iv) What do you make of the fact that longer tenure with the company, holding the other factors
fixed, is associated with a lower salary?
12 The following analysis was obtained using data in MEAP93, which contains school-level pass rates
(as a percent) on a tenth-grade math test.
(i) The variable expend is expenditures per student, in dollars, and math10 is the pass rate on the
exam. The following simple regression relates math10 to lexpend = log(expend):
math10 5 269.34 1 11.16 lexpend
1 25.53 2
2
1 3.17 2
n 5 408, R 5 .0297.
Interpret the coefficient on lexpend. In particular, if expend increases by 10%, what is the estimated percentage point change in math10? What do you make of the large negative intercept
estimate? (The minimum value of lexpend is 8.11 and its average value is 8.37.)
(ii) Does the small R-squared in part (i) imply that spending is correlated with other factors
affecting math10? Explain. Would you expect the R-squared to be much higher if expenditures were randomly assigned to schools—that is, independent of other school and student
­characteristics—rather than having the school districts determine spending?
(iii) When log of enrollment and the percent of students eligible for the federal free lunch program
are included, the estimated equation becomes
math10 5 223.14 1 7.75 lexpend 2 1.26 lenroll 2 .324 lnchprg
1 24.99 2 1 3.04 2
2
n 5 408, R 5 .1893.
(iv)
1 0.58 2
1 0.36 2
Comment on what happens to the coefficient on lexpend. Is the spending coefficient still statistically different from zero?
What do you make of the R-squared in part (iii)? What are some other factors that could be used
to explain math10 (at the school level)?
13 The data in MEAPSINGLE were used to estimate the following equations relating school-level p­ erformance
on a fourth-grade math test to socioeconomic characteristics of students attending school. The variable free,
58860_ch04_hr_117-162.indd 158
10/19/18 8:33 PM
CHAPTER 4 Multiple Regression Analysis: Inference
159
measured at the school level, is the percentage of students eligible for the federal free lunch program. The
variable medinc is median income in the ZIP code, and pctsgle is percent of students not living with two
parents (also measured at the ZIP code level). See also Computer Exercise C11 in Chapter 3.
math4 5 96.77 2 .833 pctsgle
1 1.60 2 1 .071 2
n 5 299, R2 5 .380
math4 5 93.00 2 .275 pctsgle 2 .402 free
1 1.63 2 1 .117 2
1 .070 2
2
n 5 299, R 5 .459
math4 5 24.49 2 .274 pctsgle 2 .422 free 2 .752 lmedinc 1 9.01 lexppp
1 59.24 2 1 .161 2
2
n 5 299, R 5 .472
1 .071 2
1 5.358 2
1 4.04 2
math4 5 17.52 2 .259 pctsgle 2 .420 free 1 8.80 lexppp
1 32.25 2 1 .117 2
2
n 5 299, R 5 .472.
(i)
1 .070 2
1 3.76 2
Interpret the coefficient on the variable pctsgle in the first equation. Comment on what happens
when free is added as an explanatory variable.
(ii) Does expenditure per pupil, entered in logarithmic form, have a statistically significant effect on
performance? How big is the estimated effect?
(iii) If you had to choose among the four equations as your best estimate of the effect of pctsgle and
obtain a 95% confidence interval of bpctsgle, which would you choose? Why?
Computer Exercises
C1 The following model can be used to study whether campaign expenditures affect election outcomes:
voteA 5 b0 1 b1log 1 expendA 2 1 b2log 1 expendB 2 1 b3prtystrA 1 u,
where voteA is the percentage of the vote received by Candidate A, expendA and expendB are campaign expenditures by Candidates A and B, and prtystrA is a measure of party strength for Candidate A
(the percentage of the most recent presidential vote that went to A’s party).
(i) What is the interpretation of b1?
(ii) In terms of the parameters, state the null hypothesis that a 1% increase in A’s expenditures is
offset by a 1% increase in B’s expenditures.
(iii) Estimate the given model using the data in VOTE1 and report the results in usual form. Do A’s
expenditures affect the outcome? What about B’s expenditures? Can you use these results to test
the hypothesis in part (ii)?
(iv) Estimate a model that directly gives the t statistic for testing the hypothesis in part (ii). What do
you conclude? (Use a two-sided alternative.)
C2 Use the data in LAWSCH85 for this exercise.
(i) Using the same model as in Problem 4 in Chapter 3, state and test the null hypothesis that the
rank of law schools has no ceteris paribus effect on median starting salary.
(ii) Are features of the incoming class of students—namely, LSAT and GPA—individually or jointly
significant for explaining salary? (Be sure to account for missing data on LSAT and GPA.)
58860_ch04_hr_117-162.indd 159
10/19/18 8:33 PM
160
PART 1 Regression Analysis with Cross-Sectional Data
(iii) Test whether the size of the entering class (clsize) or the size of the faculty (faculty) needs to be
added to this equation; carry out a single test. (Be careful to account for missing data on clsize
and faculty.)
(iv) What factors might influence the rank of the law school that are not included in the salary regression?
C3 Refer to Computer Exercise C2 in Chapter 3. Now, use the log of the housing price as the dependent
variable:
(i)
log 1 price 2 5 b0 1 b1sqrft 1 b2bdrms 1 u.
You are interested in estimating and obtaining a confidence interval for the percentage
change in price when a 150-square-foot bedroom is added to a house. In decimal form, this is
u 1 5 150b1 1 b2. Use the data in HPRICE1 to estimate u 1.
(ii) Write b2 in terms of u 1 and b1 and plug this into the log(price) equation.
(iii) Use part (ii) to obtain a standard error for u^ 1 and use this standard error to construct a 95%
confidence interval.
C4 In Example 4.9, the restricted version of the model can be estimated using all 1,388 observations in
the sample. Compute the R-squared from the regression of bwght on cigs, parity, and faminc using all
observations. Compare this to the R-squared reported for the restricted model in Example 4.9.
C5 Use the data in MLB1 for this exercise.
(i) Use the model estimated in equation (4.31) and drop the variable rbisyr. What happens to the
statistical significance of hrunsyr? What about the size of the coefficient on hrunsyr?
(ii) Add the variables runsyr (runs per year), fldperc (fielding percentage), and sbasesyr
­(stolen bases per year) to the model from part (i). Which of these factors are individually
significant?
(iii) In the model from part (ii), test the joint significance of bavg, fldperc, and sbasesyr.
C6 Use the data in WAGE2 for this exercise.
(i) Consider the standard wage equation
log 1 wage 2 5 b0 1 b1educ 1 b2exper 1 b3tenure 1 u.
(ii)
State the null hypothesis that another year of general workforce experience has the same effect
on log(wage) as another year of tenure with the current employer.
Test the null hypothesis in part (i) against a two-sided alternative, at the 5% significance level,
by constructing a 95% confidence interval. What do you conclude?
C7 Refer to the example used in Section 4-4. You will use the data set TWOYEAR.
(i) The variable phsrank is the person’s high school percentile. (A higher number is better. For
example, 90 means you are ranked better than 90 percent of your graduating class.) Find the
smallest, largest, and average phsrank in the sample.
(ii) Add phsrank to equation (4.26) and report the OLS estimates in the usual form. Is phsrank
statistically significant? How much is 10 percentage points of high school rank worth in terms
of wage?
(iii) Does adding phsrank to (4.26) substantively change the conclusions on the returns to twoand four-year colleges? Explain.
(iv) The data set contains a variable called id. Explain why if you add id to equation (4.17) or (4.26)
you expect it to be statistically insignificant. What is the two-sided p-value?
C8 The data set 401KSUBS contains information on net financial wealth (nettfa), age of the survey
respondent (age), annual family income (inc), family size (fsize), and participation in certain pension
plans for people in the United States. The wealth and income variables are both recorded in thousands
of dollars. For this question, use only the data for single-person households (so fsize 5 1).
(i) How many single-person households are there in the data set?
58860_ch04_hr_117-162.indd 160
10/19/18 8:33 PM
CHAPTER 4 Multiple Regression Analysis: Inference
(ii)
161
Use OLS to estimate the model
nettfa 5 b0 1 b1inc 1 b2age 1 u,
and report the results using the usual format. Be sure to use only the single-person households in the sample. Interpret the slope coefficients. Are there any surprises in the slope
estimates?
(iii) Does the intercept from the regression in part (ii) have an interesting meaning? Explain.
(iv) Find the p-value for the test H0: b2 5 1 against H1: b2 , 1. Do you reject H0 at the 1%
significance level?
(v) If you do a simple regression of nettfa on inc, is the estimated coefficient on inc much different
from the estimate in part (ii)? Why or why not?
C9 Use the data in DISCRIM to answer this question. (See also Computer Exercise C8 in Chapter 3.)
(i) Use OLS to estimate the model
log 1 psoda 2 5 b0 1 b1prpblck 1 b2log 1 income 2 1 b3prppov 1 u,
and report the results in the usual form. Is b^ 1 statistically different from zero at the 5% level
against a two-sided alternative? What about at the 1% level?
(ii) What is the correlation between log(income) and prppov? Is each variable statistically significant in any case? Report the two-sided p-values.
(iii) To the regression in part (i), add the variable log(hseval). Interpret its coefficient and report the
two-sided p-value for H0: blog1hseval2 5 0.
(iv) In the regression in part (iii), what happens to the individual statistical significance of
log(income) and prppov? Are these variables jointly significant? (Compute a p-value.) What do
you make of your answers?
(v) Given the results of the previous regressions, which one would you report as most reliable in
determining whether the racial makeup of a zip code influences local fast-food prices?
C10 Use the data in ELEM94_95 to answer this question. The findings can be compared with those in
Table 4.1. The dependent variable lavgsal is the log of average teacher salary and bs is the ratio of
average benefits to average salary (by school).
(i) Run the simple regression of lavgsal on bs. Is the estimated slope statistically different from
zero? Is it statistically different from 21?
(ii) Add the variables lenrol and lstaff to the regression from part (i). What happens to the coefficient on bs? How does the situation compare with that in Table 4.1?
(iii) Why is the standard error on the bs coefficient smaller in part (ii) than in part (i)? (Hint: What
happens to the error variance versus multicollinearity when lenrol and lstaff are added?)
(iv) How come the coefficient on lstaff is negative? Is it large in magnitude?
(v) Now add the variable lunch to the regression. Holding other factors fixed, are teachers being
compensated for teaching students from disadvantaged backgrounds? Explain.
(vi) Overall, is the pattern of results that you find with ELEM94_95 consistent with the pattern
in Table 4.1?
C11 Use the data in HTV to answer this question. See also Computer Exercise C10 in Chapter 3.
(i) Estimate the regression model
educ 5 b0 1 b1motheduc 1 b2fatheduc 1 b3abil 1 b4abil2 1 u
by OLS and report the results in the usual form. Test the null hypothesis that educ is linearly
related to abil against the alternative that the relationship is quadratic.
(ii) Using the equation in part (i), test H0: b1 5 b2 against a two-sided alternative. What is the
p-value of the test?
(iii) Add the two college tuition variables to the regression from part (i) and determine whether they
are jointly statistically significant.
58860_ch04_hr_117-162.indd 161
10/19/18 8:33 PM
162
PART 1 Regression Analysis with Cross-Sectional Data
(iv)
(v)
What is the correlation between tuit17 and tuit18? Explain why using the average of the tuition
over the two years might be preferred to adding each separately. What happens when you do use
the average?
Do the findings for the average tuition variable in part (iv) make sense when interpreted
causally? What might be going on?
C12 Use the data in ECONMATH to answer the following questions.
(i) Estimate a model explaining colgpa to hsgpa, actmth, and acteng. Report the results in the usual
form. Are all explanatory variables statistically significant?
(ii) Consider an increase in hsgpa of one standard deviation, about .343. By how much does colgpa
increase, holding actmth and acteng fixed. About how many standard deviations would the actmth have to increase to change colgpa by the same amount as one standard deviation in hsgpa?
Comment.
(iii) Test the null hypothesis that actmth and acteng have the same effect (in the population) against
a two-sided alternative. Report the p-value and describe your conclusions.
(iv) Suppose the college admissions officer wants you to use the data on the variables in part (i) to
create an equation that explains at least 50 percent of the variation in colgpa. What would you
tell the officer?
C13 Use the data set GPA1 to answer this question. It was used in Computer Exercise C13 in Chapter 3 to
estimate the effect of PC ownership on college GPA.
(i) Run the regression colGPA on PC, hsGPA, and ACT and obtain a 95% confidence interval
for bPC. Is the estimated coefficient statistically significant at the 5% level against a two-sided
alternative?
(ii) Discuss the statistical significance of the estimates b^ hsGPA and b^ ACT in part (i). Is hsGPA or ACT
the more important predictor of colGPA?
(iii) Add the two indicators fathcoll and mothcoll to the regression in part (i). Is either
individually significant? Are they jointly statistically significant at the 5% level?
C14 Use the data set JTRAIN98 to answer this question.
(i) Add the unemployment indicator unem96 to the regression reported in equation (4.52). Interpret
its coefficient and discuss whether its sign and magnitude seem sensible. Is the estimate statistically significant?
(ii) What happens to the estimated job training effect compared with equation (4.52)? Is it still
economically and statistically significant?
(iii) Find the correlation between earn96 and unem96. Is it about what you would expect? Explain.
(iv) Do you think your finding in part (iii) means you should drop unem96 from the regression?
Explain.
58860_ch04_hr_117-162.indd 162
10/19/18 8:33 PM
chapter
5
Multiple Regression
Analysis: OLS
Asymptotics
I
n Chapters 3 and 4, we covered what are called finite sample, small sample, or exact properties of
the OLS estimators in the population model
y 5 b0 1 b1x1 1 b2x2 1 p 1 bkxk 1 u.
[5.1]
For example, the unbiasedness of OLS (derived in Chapter 3) under the first four Gauss-Markov
assumptions is a finite sample property because it holds for any sample size n (subject to the mild
restriction that n must be at least as large as the total number of parameters in the regression model,
k 1 1). Similarly, the fact that OLS is the best linear unbiased estimator under the full set of GaussMarkov assumptions (MLR.1 through MLR.5) is a finite sample property.
In Chapter 4, we added the classical linear model Assumption MLR.6, which states that the error
term u is normally distributed and independent of the explanatory variables. This allowed us to derive
the exact sampling distributions of the OLS estimators (conditional on the explanatory variables in
the sample). In particular, Theorem 4.1 showed that the OLS estimators have normal sampling distributions, which led directly to the t and F distributions for t and F statistics. If the error is not normally
distributed, the distribution of a t statistic is not exactly t, and an F statistic does not have an exact
F distribution for any sample size.
In addition to finite sample properties, it is important to know the asymptotic properties or large
sample properties of estimators and test statistics. These properties are not defined for a particular
sample size; rather, they are defined as the sample size grows without bound. Fortunately, under the
assumptions we have made, OLS has satisfactory large sample properties. One practically important
163
58860_ch05_hr_163-180.indd 163
10/18/18 4:25 PM
164
PART 1 Regression Analysis with Cross-Sectional Data
finding is that even without the normality assumption (Assumption MLR.6), t and F statistics have
approximately t and F distributions, at least in large sample sizes. We discuss this in more detail in
Section 5-2, after we cover the consistency of OLS in Section 5-1.
Because the material in this chapter is more difficult to understand, and because one can conduct
empirical work without a deep understanding of its contents, this chapter may be skipped. However,
we will necessarily refer to large sample properties of OLS when we study discrete response variables
in Chapter 7, relax the homoskedasticity assumption in Chapter 8, and delve into estimation with time
series data in Part 2. Furthermore, virtually all advanced econometric methods derive their justification using large-sample analysis, so readers who will continue into Part 3 should be familiar with the
contents of this chapter.
5-1 Consistency
Unbiasedness of estimators, although important, cannot always be obtained. For example, as we discussed in Chapter 3, the standard error of the regression, s^ , is not an unbiased estimator for s, the
standard deviation of the error u, in a multiple regression model. Although the OLS estimators are
unbiased under MLR.1 through MLR.4, in Chapter 11 we will find that there are time series regressions where the OLS estimators are not unbiased. Further, in Part 3 of the text, we encounter several
other estimators that are biased yet useful.
Although not all useful estimators are unbiased, virtually all economists agree that consistency
is a minimal requirement for an estimator. The Nobel Prize–winning econometrician Clive W. J.
Granger once remarked, “If you can’t get it right as n goes to infinity, you shouldn’t be in this business.” The implication is that, if your estimator of a particular population parameter is not consistent,
then you are wasting your time.
There are a few different ways to describe consistency. Formal definitions and results are given in
Math Refresher C; here, we focus on an intuitive understanding. For concreteness, let b^ j be the OLS
estimator of bj for some j. For each n, b^ j has a probability distribution (representing its possible values in different random samples of size n). Because b^ j is unbiased under Assumptions MLR.1 through
MLR.4, this distribution has mean value bj. If this estimator is consistent, then the distribution of b^ j
becomes more and more tightly distributed around bj as the sample size grows. As n tends to infinity,
the distribution of b^ j collapses to the single point bj. In effect, this means that we can make our estimator arbitrarily close to bj if we can collect as much data as we want. This convergence is illustrated
in Figure 5.1.
Naturally, for any application, we have a fixed sample size, which is a major reason an asymptotic property such as consistency can be difficult to grasp. Consistency involves a thought experiment about what would happen as the sample size gets large (while, at the same time, we obtain
numerous random samples for each sample size). If obtaining more and more data does not generally
get us closer to the parameter value of interest, then we are using a poor estimation procedure.
Conveniently, the same set of assumptions implies both unbiasedness and consistency of OLS.
We summarize with a theorem.
Theorem
5.1
Consistency of OLS
Under Assumptions MLR.1 through MLR.4, the OLS estimator b^ j is consistent for bj , for all
j 5 0, 1, . . . , k.
58860_ch05_hr_163-180.indd 164
10/18/18 4:25 PM
CHAPTER 5 Multiple Regression Analysis: OLS Asymptotics
165
Figure 5.1 Sampling distributions of b^ 1 for sample sizes n1 , n2 , n3.
n3
fˆ
n2
n1
ˆ1
1
A general proof of this result is most easily developed using the matrix algebra methods described
in Appendices D and E. But we can prove Theorem 5.1 without difficulty in the case of the simple
regression model. We focus on the slope estimator, b^ 1.
The proof starts out the same as the proof of unbiasedness: we write down the formula for b^ 1, and
then plug in yi 5 b0 1 b1xi1 1 ui:
b^ 1 5 a a 1 xi1 2 x1 2 yi b ^ a a 1 xi1 2 x1 2 2 b
n
n
i51
i51
5 b1 1 an21 a 1 xi1 2 x1 2 ui b ^ an21 a 1 xi1 2 x1 2 2 b,
n
n
i51
i51
[5.2]
where dividing both the numerator and denominator by n does not change the expression
but allows us to directly apply the law of large numbers. When we apply the law of large
numbers to the averages in the second part of equation (5.2), we conclude that the numerator
and denominator converge in probability to the population quantities, Cov 1 x1,u 2 and Var 1 x1 2 ,
­respectively. Provided that Var 1 x1 2 2 0—which is assumed in MLR.3—we can use the properties of
probability limits (see Math Refresher C) to get
plim b^ 1 5 b1 1 Cov 1 x1,u 2 /Var 1 x1 2
5 b1 because Cov 1 x1,u 2 5 0.
[5.3]
We have used the fact, discussed in Chapters 2 and 3, that E 1 u 0 x1 2 5 0 (Assumption MLR.4) implies
that x1 and u are uncorrelated (have zero covariance).
58860_ch05_hr_163-180.indd 165
10/18/18 4:25 PM
166
PART 1 Regression Analysis with Cross-Sectional Data
As a technical matter, to ensure that the probability limits exist, we should assume that
Var 1 x1 2 , ` and Var(u) , ` (which means that their probability distributions are not too spread
out), but we will not worry about cases where these assumptions might fail. Further, we could—and,
in an advanced treatment of econometrics, we would—explicitly relax Assumption MLR.3 to rule
out only perfect collinearity in the population. As stated, Assumption MLR.3 also disallows perfect collinearity among the regressors in the sample we have at hand. Technically, for the thought
experiment we can show consistency with no perfect collinearity in the population, allowing for the
unlucky possibility that we draw a data set that does exhibit perfect collinearity. From a practical
perspective the distinction is unimportant, as we cannot compute the OLS estimates for our sample
if MLR.3 fails.
The previous arguments, and equation (5.3) in particular, show that OLS is consistent in the simple regression case if we assume only zero correlation. This is also true in the general case. We now
state this as an assumption.
Assumption MLR.49
Zero Mean and Zero Correlation
E 1 u 2 5 0 and Cov 1 xj,u 2 5 0, for j 5 1, 2, . . . , k.
Assumption MLR.4 r is weaker than Assumption MLR.4 in the sense that the latter implies the
former. One way to characterize the zero conditional mean assumption, E 1 u 0 x1, . . . , xk 2 5 0, is that
any function of the explanatory variables is uncorrelated with u. Assumption MLR.4 r requires only
that each xj is uncorrelated with u (and that u has a zero mean in the population). In Chapter 2, we
actually motivated the OLS estimator for simple regression using Assumption MLR.4 r, and the first
order conditions for OLS in the multiple regression case, given in equation (3.13), are simply the sample analogs of the population zero correlation assumptions (and zero mean assumption). Therefore, in
some ways, Assumption MLR.4 r is more natural an assumption because it leads directly to the OLS
estimates. Further, when we think about violations of Assumption MLR.4, we usually think in terms
of Cov 1 xj,u 2 2 0 for some j. So how come we have used Assumption MLR.4 until now? There are
two reasons, both of which we have touched on earlier. First, OLS turns out to be biased (but consistent) under Assumption MLR.4r if E 1 u 0 x1, . . . , xk 2 depends on any of the xj. Because we have previously focused on finite sample, or exact, sampling properties of the OLS estimators, we have needed
the stronger zero conditional mean assumption.
Second, and probably more important, is that the zero conditional mean assumption means that
we have properly modeled the population regression function (PRF). That is, under Assumption
MLR.4 we can write
E 1 y 0 x1, . . . , xk 2 5 b0 1 b1x1 1 p 1 bkxk,
and so we can obtain partial effects of the explanatory variables on the average or expected value
of y. If we instead only assume Assumption MLR.4r, b0 1 b1x1 1 p 1 bkxk need not represent the
PRF, and we face the possibility that some nonlinear functions of the xj, such as x2j , could be correlated with the error u. A situation like this means that we have neglected nonlinearities in the model
that could help us better explain y; if we knew that, we would usually include such nonlinear functions. In other words, most of the time we hope to get a good estimate of the PRF, and so the zero
conditional mean assumption is natural. Nevertheless, the weaker zero correlation assumption turns
out to be useful in interpreting OLS estimation of a linear model as providing the best linear approximation to the PRF. It is also used in more advanced settings, such as in Chapter 15, where we have
no interest in modeling a PRF. For further discussion of this somewhat subtle point, see Wooldridge
(2010, Chapter 4) as well as Problem 6 at the end of this chapter.
58860_ch05_hr_163-180.indd 166
10/18/18 4:25 PM
CHAPTER 5 Multiple Regression Analysis: OLS Asymptotics
167
5-1a Deriving the Inconsistency in OLS
Just as failure of E 1 u 0 x1, . . . , xk 2 5 0 causes bias in the OLS estimators, correlation between u and
any of x1, x2, . . . , xk generally causes all of the OLS estimators to be inconsistent. This simple but
important observation is often summarized as: if the error is correlated with any of the independent
variables, then OLS is biased and inconsistent. This is very unfortunate because it means that any bias
persists as the sample size grows.
In the simple regression case, we can obtain the inconsistency from the first part of equation (5.3),
which holds whether or not u and x1 are uncorrelated. The inconsistency in b^ 1 (sometimes loosely
called the asymptotic bias) is
plim b^ 1 2 b1 5 Cov 1 x1,u 2 /Var 1 x1 2 .
[5.4]
Because Var 1 x1 2 . 0, the inconsistency in b^ 1 is positive if x1 and u are positively correlated, and the
inconsistency is negative if x1 and u are negatively correlated. If the covariance between x1 and u is
small relative to the variance in x1, the inconsistency can be negligible; unfortunately, we cannot even
estimate how big the covariance is because u is unobserved.
We can use (5.4) to derive the asymptotic analog of the omitted variable bias (see Table 3.2 in
Chapter 3). Suppose the true model,
y 5 b0 1 b1x1 1 b2x2 1 v,
satisfies the first four Gauss-Markov assumptions. Then v has a zero mean and is uncorrelated with
x1 and x2. If b^ 0, b^ 1, and b^ 2 denote the OLS estimators from the regression of y on x1 and x2, then
Theorem 5.1 implies that these estimators are consistent. If we omit x2 from the regression and do the
| denote the simple regression slope estimator.
simple regression of y on x1, then u 5 b2x2 1 v. Let b
1
Then
| 5b 1bd,
plim b
[5.5]
1
1
2 1
where
d1 5 Cov 1 x1,x2 2 /Var 1 x1 2 .
[5.6]
Thus, for practical purposes, we can view the inconsistency as being the same as the bias. The difference is that the inconsistency is expressed in terms of the population variance of x1 and the population
covariance between x1 and x2, while the bias is based on their sample counterparts (because we condition on the values of x1 and x2 in the sample).
| is a consistent estimator of
If x1 and x2 are uncorrelated (in the population), then d1 5 0, and b
1
b1 (although not necessarily unbiased). If x2 has a positive partial effect on y, so that b2 . 0, and x1
| is positive, and so on.
and x2 are positively correlated, so that d1 . 0, then the inconsistency in b
1
We can obtain the direction of the inconsistency or asymptotic bias from Table 3.2. If the covariance
between x1 and x2 is small relative to the variance of x1, the inconsistency can be small.
Example 5.1
Housing Prices and Distance from an Incinerator
Let y denote the price of a house (price), let x1 denote the distance from the house to a new trash
­incinerator (distance), and let x2 denote the “quality” of the house (quality). The variable quality is
left vague so that it can include things like size of the house and lot, number of bedrooms and bathrooms, and intangibles such as attractiveness of the neighborhood. If the incinerator depresses house
prices, then b1 should be positive: everything else being equal, a house that is farther away from the
incinerator is worth more. By definition, b2 is positive because higher quality houses sell for more,
other factors being equal. If the incinerator was built farther away, on average, from better homes, then
­distance and quality are positively correlated, and so d1 . 0. A simple regression of price on distance
[or log(price) on log(distance)] will tend to overestimate the effect of the incinerator: b1 1 b2 d1 . b1.
58860_ch05_hr_163-180.indd 167
10/18/18 4:25 PM
168
PART 1 Regression Analysis with Cross-Sectional Data
G o i n g F u rt h e r 5 . 1
Suppose that the model
score 5 b0 1 b1skipped 1 b2 priGPA 1 u
satisfies the first four Gauss-Markov
­a ssumptions, where score is score on a
final exam, skipped is number of classes
skipped, and priGPA is GPA prior to the
| is from the simple
current semester. If b
1
regression of score on skipped, what is the
|?
direction of the asymptotic bias in b
1
An important point about inconsistency in OLS
estimators is that, by definition, the problem does
not go away by adding more observations to the
sample. If anything, the problem gets worse with
more data: the OLS estimator gets closer and closer
to b1 1 b2d1 as the sample size grows.
Deriving the sign and magnitude of the inconsistency in the general k regressor case is harder,
just as deriving the bias is more difficult. We need
to remember that if we have the model in equation
(5.1) where, say, x1 is correlated with u but the other
independent variables are uncorrelated with u, all of
the OLS estimators are generally inconsistent. For
example, in the k 5 2 case,
y 5 b0 1 b1x1 1 b2x2 1 u,
suppose that x2 and u are uncorrelated but x1 and u are correlated. Then the OLS estimators b^ 1 and b^ 2
will generally both be inconsistent. (The intercept will also be inconsistent.) The inconsistency in b^ 2
arises when x1 and x2 are correlated, as is usually the case. If x1 and x2 are uncorrelated, then any correlation between x1 and u does not result in the inconsistency of b^ 2: plim b^ 2 5 b2. Further, the inconsistency in b^ 1 is the same as in (5.4). The same statement holds in the general case: if x1 is correlated
with u, but x1 and u are uncorrelated with the other independent variables, then only b^ 1 is inconsistent,
and the inconsistency is given by (5.4). The general case is very similar to the omitted variable case in
Section 3A.4 of Appendix 3A.
5-2 Asymptotic Normality and Large Sample Inference
Consistency of an estimator is an important property, but it alone does not allow us to perform
statistical inference. Simply knowing that the estimator is getting closer to the population value
as the sample size grows does not allow us to test hypotheses about the parameters. For testing, we need the sampling distribution of the OLS estimators. Under the classical linear model
assumptions MLR.1 through MLR.6, Theorem 4.1 shows that the sampling distributions are normal. This result is the basis for deriving the t and F distributions that we use so often in applied
econometrics.
The exact normality of the OLS estimators hinges crucially on the normality of the distribution
of the error, u, in the population. If the errors u1, u2, . . . , un are random draws from some distribution
other than the normal, the b^ j will not be normally distributed, which means that the t statistics will
not have t distributions and the F statistics will not have F distributions. This is a potentially serious
problem because our inference hinges on being able to obtain critical values or p-values from the t or
F distributions.
Recall that Assumption MLR.6 is equivalent to saying that the distribution of y given
x1, x2, . . . , xk is normal. Because y is observed and u is not, in a particular application, it is much
easier to think about whether the distribution of y is likely to be normal. In fact, we have already
seen a few examples where y definitely cannot have a conditional normal distribution. A normally
distributed random variable is symmetrically distributed about its mean, it can take on any positive or negative value, and more than 95% of the area under the distribution is within two standard
deviations.
58860_ch05_hr_163-180.indd 168
10/18/18 4:25 PM
CHAPTER 5 Multiple Regression Analysis: OLS Asymptotics
169
Figure 5.2 Histogram of prate using the data in 401K.
proportion in cell
.8
.6
.4
.2
0
0
10
20
30
40
50
60
70
80
90
100
participation rate (in percentage form)
In Example 3.5, we estimated a model explaining the number of arrests of young men during a particular year (narr86). In the population, most men are not arrested during the year, and
the vast majority are arrested one time at the most. (In the sample of 2,725 men in the data set
CRIME1, fewer than 8% were arrested more than once during 1986.) Because narr86 takes on
only two values for 92% of the sample, it cannot be close to being normally distributed in the
population.
In Example 4.6, we estimated a model explaining participation percentages (prate) in 401(k)
pension plans. The frequency distribution (also called a histogram) in Figure 5.2 shows that the distribution of prate is heavily skewed to the right, rather than being normally distributed. In fact, over
40% of the observations on prate are at the value 100, indicating 100% participation. This violates the
normality assumption even conditional on the explanatory variables.
We know that normality plays no role in the unbiasedness of OLS, nor does it affect the conclusion that OLS is the best linear unbiased estimator under the Gauss-Markov assumptions. But exact
inference based on t and F statistics requires MLR.6. Does this mean that, in our prior analysis of
prate in Example 4.6, we must abandon the t statistics for determining which variables are statistically significant? Fortunately, the answer to this question is no. Even though the yi are not from a
normal distribution, we can use the central limit theorem from Math Refresher C to conclude that the
OLS estimators satisfy asymptotic normality, which means they are approximately normally distributed in large enough sample sizes.
58860_ch05_hr_163-180.indd 169
10/18/18 4:25 PM
170
PART 1 Regression Analysis with Cross-Sectional Data
Theorem
5.2
Asymptotic Normality of OLS
Under the Gauss-Markov Assumptions MLR.1 through MLR.5,
a
(i) !n 1 b^ j 2 bj 2 | Normal 1 0, s2 /a2j 2 , where s2 /a2j . 0 is the asymptotic variance of !n 1 b^ j 2 bj 2 ;
n
for the slope coefficients, a2j 5 plim 1 n21 g i51 r^ij2 2 , where the r^ij are the residuals from regressing xj on the
other independent variables. We say that b^ j is asymptotically normally distributed (see Math Refresher C);
(ii) s^ 2 is a consistent estimator of s2 5 Var 1 u 2 ;
(iii) For each j,
a
and
1 b^ j 2 bj 2 /sd 1 b^ j 2 | Normal 1 0,1 2
a
1 b^ j 2 bj 2 /se 1 b^ j 2 | Normal 1 0,1 2 ,
[5.7]
where se 1 b^ j 2 is the usual OLS standard error.
The proof of asymptotic normality is somewhat complicated and is sketched in the appendix for
the simple regression case. Part (ii) follows from the law of large numbers, and part (iii) follows from
parts (i) and (ii) and the asymptotic properties discussed in Math Refresher C.
Theorem 5.2 is useful because the normality Assumption MLR.6 has been dropped; the only
restriction on the distribution of the error is that it has finite variance, something we will always
assume. We have also assumed zero conditional mean (MLR.4) and homoskedasticity of u (MLR.5).
In trying to understand the meaning of Theorem 5.2, it is important to keep separate the notions
of the population distribution of the error term, u, and the sampling distributions of the b^ j as the
sample size grows. A common mistake is to think that something is happening to the distribution of
u—namely, that it is getting “closer” to normal—as the sample size grows. But remember that the
population distribution is immutable and has nothing to do with the sample size. For example, we
previously discussed narr86, the number of times a young man is arrested during the year 1986. The
nature of this variable—it takes on small, nonnegative integer values—is fixed in the population.
Whether we sample 10 men or 1,000 men from this population obviously has no effect on the population distribution.
What Theorem 5.2 says is that, regardless of the population distribution of u, the OLS estimators, when properly standardized, have approximate standard normal distributions. This approximation comes about by the central limit theorem because the OLS estimators involve—in a complicated
way—the use of sample averages. Effectively, the sequence of distributions of averages of the underlying errors is approaching normality for virtually any population distribution.
Notice how the standardized b^ j has an asymptotic standard normal distribution whether we divide
the difference b^ j 2 bj by sd 1 b^ j 2 (which we do not observe because it depends on s) or by se 1 b^ j 2
(which we can compute from our data because it depends on s^ ). In other words, from an asymptotic
point of view it does not matter that we have to replace s with s^ . Of course, replacing s with s^ affects
the exact distribution of the standardized b^ j. We just saw in Chapter 4 that under the classical linear
model assumptions, 1 b^ j 2 bj 2 /sd 1 b^ j 2 has an exact Normal(0,1) distribution and 1 b^ j 2 bj 2 /se 1 b^ j 2 has
an exact tn2k21 distribution.
How should we use the result in equation (5.7)? It may seem one consequence is that, if we are
going to appeal to large-sample analysis, we should now use the standard normal distribution for
inference rather than the t distribution. But from a practical perspective it is just as legitimate to write
58860_ch05_hr_163-180.indd 170
10/18/18 4:25 PM
CHAPTER 5 Multiple Regression Analysis: OLS Asymptotics
1 b^ j 2 bj 2 /se 1 b^ j 2 |a tn2k21 5 tdf,
171
[5.8]
because tdf approaches the Normal(0,1) distribution as df gets large. Because we know under the
CLM assumptions the tn2k21 distribution holds exactly, it makes sense to treat 1 b^ j 2 bj 2 /se 1 b^ j 2 as a
tn2k21 random variable generally, even when MLR.6 does not hold.
Equation (5.8) tells us that t testing and the construction of confidence intervals are carried out
exactly as under the classical linear model assumptions. This means that our analysis of dependent
variables like prate and narr86 does not have to change at all if the Gauss-Markov assumptions hold:
in both cases, we have at least 1,500 observations, which is certainly enough to justify the approximation of the central limit theorem.
If the sample size is not very large, then the t distribution can be a poor approximation to the
distribution of the t statistics when u is not normally distributed. Unfortunately, there are no general
prescriptions on how big the sample size must be before the approximation is good enough. Some
econometricians think that n 5 30 is satisfactory, but this cannot be sufficient for all possible distributions of u. Depending on the distribution of u, more observations may be necessary before the central
limit theorem delivers a useful approximation. Further, the quality of the approximation depends not
just on n, but on the df, n 2 k 2 1: with more independent variables in the model, a larger sample
size is usually needed to use the t approximation. Methods for inference with small degrees of freedom and nonnormal errors are outside the scope of this text. We will simply use the t statistics as we
always have without worrying about the normality assumption.
It is very important to see that Theorem 5.2 does require the homoskedasticity assumption (along
with the zero conditional mean assumption). If Var(y|x) is not constant, the usual t statistics and confidence intervals are invalid no matter how large the sample size is; the central limit theorem does not
bail us out when it comes to heteroskedasticity. For this reason, we devote all of Chapter 8 to discussing what can be done in the presence of heteroskedasticity.
One conclusion of Theorem 5.2 is that s^ 2 is a consistent estimator of s2; we already know
from Theorem 3.3 that s^ 2 is unbiased for s2 under the Gauss-Markov assumptions. The consistency
implies that s^ is a consistent estimator of s, which is important in establishing the asymptotic normality result in equation (5.7).
Remember that s^ appears in the standard error for each b^ j. In fact, the estimated variance of b^ j is
Var 1 b^ j 2 5
s^ 2
,
SSTj 1 1 2 R2j 2
[5.9]
where SSTj is the total sum of squares of xj in the
sample, and R2j is the R-squared from regressing xj
G o i n g F u rt h e r 5 . 2
on all of the other independent variables. In Section
In a regression model with a large sample
3-4, we studied each component of (5.9), which we
size, what is an approximate 95% confiwill now expound on in the context of asymptotic
dence interval for b^ j under MLR.1 through
analysis. As the sample size grows, s^ 2 converges in
MLR.5? We call this an asymptotic confiprobability to the constant s2. Further, R2j approaches
dence interval.
a number strictly between zero and unity (so that
1 2 R2j converges to some number between zero and
one). The sample variance of xj is SSTj/n, and so SSTj/n converges to Var 1 xj 2 as the sample size
grows. This means that SSTj grows at approximately the same rate as the sample size: SSTj < ns2j ,
where s2j is the population variance of xj. When we combine these facts, we find that Var 1 b^ j 2 shrinks
to zero at the rate of 1/n; this is why larger sample sizes are better.
When u is not normally distributed, the square root of (5.9) is sometimes called the asymptotic
standard error, and t statistics are called asymptotic t statistics. Because these are the same quantities we dealt with in Chapter 4, we will just call them standard errors and t statistics, with the understanding that sometimes they have only large-sample justification. A similar comment holds for an
asymptotic confidence interval constructed from the asymptotic standard error.
58860_ch05_hr_163-180.indd 171
10/18/18 4:25 PM
172
PART 1 Regression Analysis with Cross-Sectional Data
Using the preceding argument about the estimated variance, we can write
se 1 b^ j 2 < cj/ !n,
[5.10]
where cj is a positive constant that does not depend on the sample size. In fact, the constant cj can be
shown to be
cj 5
s
sj"1 2 r2j
,
where s 5 sd 1 u 2 , sj 5 sd 1 xj 2 , and r2j is the population R-squared from regressing xj on the other
explanatory variables. Just like studying equation (5.9) to see which variables affect Var 1 b^ j 2 under
the Gauss-Markov assumptions, we can use this expression for cj to study the impact of larger
error standard deviation 1 s 2 , more population variation in xj 1 sj 2 , and multicollinearity in the population 1 r2j 2 .
Equation (5.10) is only an approximation, but it is a useful rule of thumb: standard errors can be
expected to shrink at a rate that is the inverse of the square root of the sample size.
Example 5.2
Standard Errors in a Birth Weight Equation
We use the data in BWGHT to estimate a relationship where log of birth weight is the dependent
variable, and cigarettes smoked per day (cigs) and log of family income are independent variables.
The total number of observations is 1,388. Using the first half of the observations (694), the standard error for b^ cigs is about .0013. The standard error using all of the observations is about .00086.
The ratio of the latter standard error to the former is .00086/.0013 < .662. This is pretty close to
!694/1,388 < .707, the ratio obtained from the approximation in (5.10). In other words, equation
(5.10) implies that the standard error using the larger sample size should be about 70.7% of the standard error using the smaller sample. This percentage is pretty close to the 66.2% we actually compute
from the ratio of the standard errors.
The asymptotic normality of the OLS estimators also implies that the F statistics have approximate F distributions in large sample sizes. Thus, for testing exclusion restrictions or other multiple
hypotheses, nothing changes from what we have done before.
5-2a Other Large Sample Tests: The Lagrange Multiplier Statistic
Once we enter the realm of asymptotic analysis, other test statistics can be used for hypothesis testing.
For most purposes, there is little reason to go beyond the usual t and F statistics: as we just saw, these
statistics have large sample justification without the normality assumption. Nevertheless, sometimes
it is useful to have other ways to test multiple exclusion restrictions, and we now cover the Lagrange
multiplier (LM) statistic, which has achieved some popularity in modern econometrics.
The name “Lagrange multiplier statistic” comes from constrained optimization, a topic beyond
the scope of this text. [See Davidson and MacKinnon (1993).] The name score statistic—which also
comes from optimization using calculus—is used as well. Fortunately, in the linear regression framework, it is simple to motivate the LM statistic without delving into complicated mathematics.
The form of the LM statistic we derive here relies on the Gauss-Markov assumptions, the same
assumptions that justify the F statistic in large samples. We do not need the normality assumption.
To derive the LM statistic, consider the usual multiple regression model with k independent
variables:
y 5 b0 1 b1x1 1 p 1 bkxk 1 u.
58860_ch05_hr_163-180.indd 172
[5.11]
10/18/18 4:25 PM
CHAPTER 5 Multiple Regression Analysis: OLS Asymptotics
173
We would like to test whether, say, the last q of these variables all have zero population parameters:
the null hypothesis is
H0: bk2q11 5 0, . . . , bk 5 0,
[5.12]
which puts q exclusion restrictions on the model (5.11). As with F testing, the alternative to (5.12) is
that at least one of the parameters is different from zero.
The LM statistic requires estimation of the restricted model only. Thus, assume that we have run
the regression
| 1b
|x 1p1b
| x 1|
y5b
u,
0
1 1
k2q k2q
[5.13]
where “~” indicates that the estimates are from the restricted model. In particular, |
u indicates the
residuals from the restricted model. (As always, this is just shorthand to indicate that we obtain the
restricted residual for each observation in the sample.)
If the omitted variables xk2q11 through xk truly have zero population coefficients, then, at least
approximately, |
u should be uncorrelated with each of these variables in the sample. This suggests
running a regression of these residuals on those independent variables excluded under H0, which is
almost what the LM test does. However, it turns out that, to get a usable test statistic, we must include
all of the independent variables in the regression. (We must include all regressors because, in general,
the omitted regressors in the restricted model are correlated with the regressors that appear in the
restricted model.) Thus, we run the regression of
|
u on x1, x2, . . . , xk.
[5.14]
This is an example of an auxiliary regression, a regression that is used to compute a test statistic but
whose coefficients are not of direct interest.
How can we use the regression output from (5.14) to test (5.12)? If (5.12) is true, the R-squared
from (5.14) should be “close” to zero, subject to sampling error, because |
u will be approximately
uncorrelated with all the independent variables. The question, as always with hypothesis testing, is
how to determine when the statistic is large enough to reject the null hypothesis at a chosen significance level. It turns out that, under the null hypothesis, the sample size multiplied by the usual
R-squared from the auxiliary regression (5.14) is distributed asymptotically as a chi-square random
variable with q degrees of freedom. This leads to a simple procedure for testing the joint significance
of a set of q independent variables.
The Lagrange Multiplier Statistic for q Exclusion
Restrictions:
Regress y on the restricted set of independent variables and save the residuals, |
u.
2
|
(ii) Regress u on all of the independent variables and obtain the R-squared, say, Ru (to distinguish it
from the R-squareds obtained with y as the dependent variable).
(iii) Compute LM 5 nR2u [the sample size times the R-squared obtained from step (ii)].
(iv) Compare LM to the appropriate critical value, c, in a x2q distribution; if LM . c, the null hypothesis is rejected. Even better, obtain the p-value as the probability that a x2q random variable exceeds
the value of the test statistic. If the p‑value is less than the desired significance level, then H0 is
rejected. If not, we fail to reject H0. The rejection rule is essentially the same as for F testing.
(i)
Because of its form, the LM statistic is sometimes referred to as the n-R-squared statistic.
Unlike with the F statistic, the degrees of freedom in the unrestricted model plays no role in carrying
out the LM test. All that matters is the number of restrictions being tested (q), the size of the auxiliary
R-squared 1 R2u 2 , and the sample size (n). The df in the unrestricted model plays no role because of the
58860_ch05_hr_163-180.indd 173
10/18/18 4:25 PM
174
PART 1 Regression Analysis with Cross-Sectional Data
asymptotic nature of the LM statistic. But we must be sure to multiply R2u by the sample size to obtain
LM; a seemingly low value of the R-squared can still lead to joint significance if n is large.
Before giving an example, a word of caution is in order. If in step (i), we mistakenly regress y
on all of the independent variables and obtain the residuals from this unrestricted regression to be
used in step (ii), we do not get an interesting statistic: the resulting R-squared will be exactly zero!
This is because OLS chooses the estimates so that the residuals are uncorrelated in samples with all
included independent variables [see equations in (3.13)]. Thus, we can only test (5.12) by regressing
the restricted residuals on all of the independent variables. (Regressing the restricted residuals on the
restricted set of independent variables will also produce R2 5 0.)
Example 5.3
Economic Model of Crime
We illustrate the LM test by using a slight extension of the crime model from Example 3.5:
narr86 5 b0 1 b1pcnv 1 b2avgsen 1 b3tottime 1 b4ptime86 1 b5qemp86 1 u,
where
narr86 5 the number of times a man was arrested.
pcnv 5 the proportion of prior arrests leading to conviction.
avgsen 5 average sentence served from past convictions.
tottime 5 total time the man has spent in prison prior to 1986 since reaching the age of 18.
ptime86 5 months spent in prison in 1986.
qemp86 5 number of quarters in 1986 during which the man was legally employed.
We use the LM statistic to test the null hypothesis that avgsen and tottime have no effect on
narr86 once the other factors have been controlled for.
In step (i), we estimate the restricted model by regressing narr86 on pcnv, ptime86, and qemp86;
the variables avgsen and tottime are excluded from this regression. We obtain the residuals |
u from
this regression, 2,725 of them. Next, we run the regression of
|
u on pcnv, ptime86, qemp86, avgsen, and tottime;
[5.15]
as always, the order in which we list the independent variables is irrelevant. This second regression
produces R2u, which turns out to be about .0015. This may seem small, but we must multiply it by n
to get the LM statistic: LM 5 2,725 1 .0015 2 < 4.09. The 10% critical value in a chi-square distribution with two degrees of freedom is about 4.61 (rounded to two decimal places; see Table G.4). Thus,
we fail to reject the null hypothesis that bavgsen 5 0 and btottime 5 0 at the 10% level. The p-value is
P 1 x22 . 4.09 2 < .129, so we would reject H0 at the 15% level.
As a comparison, the F test for joint significance of avgsen and tottime yields a p‑value of about
.131, which is pretty close to that obtained using the LM statistic. This is not surprising because,
asymptotically, the two statistics have the same probability of Type I error. (That is, they reject the
null hypothesis with the same frequency when the null is true.)
As the previous example suggests, with a large sample, we rarely see important discrepancies
between the outcomes of LM and F tests. We will use the F statistic for the most part because it is
computed routinely by most regression packages. But you should be aware of the LM statistic as it is
used in applied work.
One final comment on the LM statistic. As with the F statistic, we must be sure to use the same
observations in steps (i) and (ii). If data are missing for some of the independent variables that are
excluded under the null hypothesis, the residuals from step (i) should be obtained from a regression
on the reduced data set.
58860_ch05_hr_163-180.indd 174
10/18/18 4:25 PM
CHAPTER 5 Multiple Regression Analysis: OLS Asymptotics
175
5-3 Asymptotic Efficiency of OLS
We know that, under the Gauss-Markov assumptions, the OLS estimators are best linear unbiased.
OLS is also asymptotically efficient among a certain class of estimators under the Gauss-Markov
assumptions. A general treatment requires matrix algebra and advanced asymptotic analysis. First, we
describe the result in the simple regression case.
In the model
[5.16]
y 5 b0 1 b1x 1 u,
u has a zero conditional mean under MLR.4: E 1 u 0 x 2 5 0. This opens up a variety of consistent estimators for b0 and b1; as usual, we focus on the slope parameter, b1. Let g(x) be any function of x; for
example, g 1 x 2 5 x2 or g 1 x 2 5 1/ 1 1 1 0 x 0 2 . Then u is uncorrelated with g(x) (see Property CE.5 in
Math Refresher B). Let zi 5 g 1 xi 2 for all observations i. Then the estimator
| 5
b
a a 1 zi 2 z 2 yi b ^ a a 1 zi 2 z 2 xi b
1
n
n
i51
i51
[5.17]
is consistent for b1, provided g(x) and x are correlated. [Remember, it is possible that g(x) and x
are uncorrelated because correlation measures linear dependence.] To see this, we can plug in
| as
yi 5 b0 1 b1xi 1 ui and write b
1
| 5 b 1 21 1 z 2 z 2 u
21
b
an a i
a 1 zi 2 z 2 xi b.
1
1
i b ^ an
n
n
i51
i51
[5.18]
Now, we can apply the law of large numbers to the numerator and denominator, which converge in
probability to Cov(z,u) and Cov(z,x), respectively. Provided that Cov 1 z,u 2 2 0—so that z and x are
correlated—we have
| 5 b 1 Cov 1 z,u 2 /Cov 1 z,x 2 5 b ,
plim b
1
1
1
because Cov 1 z,u 2 5 0 under MLR.4.
| is asymptotically normal. Nevertheless, using arguments simiIt is more difficult to show that b
1
| 2 b 2 is asymptotically normal with mean
lar to those in the appendix, it can be shown that !n 1 b
1
1
2
2
zero and asymptotic variance s Var 1 z 2 / 3 Cov 1 z,x 2 4 . The asymptotic variance of the OLS estimator
is obtained when z 5 x, in which case, Cov 1 z,x 2 5 Cov 1 x,x 2 5 Var 1 x 2 . Therefore, the asymptotic
variance of !n 1 b^ 1 2 b1 2 , where b^ 1 is the OLS estimator, is s2Var 1 x 2 / 3 Var 1 x 2 4 2 5 s2 /Var 1 x 2 .Now,
the Cauchy-Schwartz inequality (see Math Refresher B.4) implies that 3 Cov 1 z,x 2 4 2 # Var 1 z 2 Var 1 x 2 ,
| 2 b 2.
which implies that the asymptotic variance of !n 1 b^ 1 2 b1 2 is no larger than that of !n 1 b
1
1
We have shown in the simple regression case that, under the Gauss-Markov assumptions, the OLS
estimator has a smaller asymptotic variance than any estimator of the form (5.17). [The estimator
in (5.17) is an example of an instrumental variables estimator, which we will study extensively in
Chapter 15.] If the homoskedasticity assumption fails, then there are estimators of the form (5.17) that
have a smaller asymptotic variance than OLS. We will see this in Chapter 8.
The general case is similar but much more difficult mathematically. In the k regressor case, the
class of consistent estimators is obtained by generalizing the OLS first order conditions:
|
|
| 2
p2b
a gj 1 xi 2 1 yi 2 b0 2 b1xi1 2
k xik 5 0, j 5 0, 1, . . . , k,
n
[5.19]
i51
where gj 1 xi 2 denotes any function of all explanatory variables for observation i. As can be seen by
comparing (5.19) with the OLS first order conditions in (3.13), we obtain the OLS estimators when
g0 1 xi 2 5 1 and gj 1 xi 2 5 xij for j 5 1, 2, . . . , k. There are infinitely many estimators that can be
defined using the equations in (5.19) because we can use any functions of the xij that we want.
58860_ch05_hr_163-180.indd 175
10/18/18 4:25 PM
176
PART 1 Regression Analysis with Cross-Sectional Data
Theorem
5.3
Asymptotic Efficiency of OLS
| denote estimators that solve equations of the form (5.19)
Under the Gauss-Markov assumptions, let b
j
and let b^ j denote the OLS estimators. Then for j 5 0, 1, 2, . . . , k, the OLS estimators have the small| 2 b 2.
est asymptotic variances: Avar!n 1 b^ j 2 bj 2 # Avar!n 1 b
j
j
Proving consistency of the estimators in (5.19), let alone showing they are asymptotically n­ ormal,
is mathematically difficult. See Wooldridge (2010, Chapter 5).
Summary
The claims underlying the material in this chapter are fairly technical, but their practical implications are
straightforward. We have shown that the first four Gauss-Markov assumptions imply that OLS is consistent. Furthermore, all of the methods of testing and constructing confidence intervals that we learned in
Chapter 4 are approximately valid without assuming that the errors are drawn from a normal distribution
(equivalently, the distribution of y given the explanatory variables is not normal). This means that we can
apply OLS and use previous methods for an array of applications where the dependent variable is not even
approximately normally distributed. We also showed that the LM statistic can be used instead of the F statistic for testing exclusion restrictions.
Before leaving this chapter, we should note that examples such as Example 5.3 may very well have
problems that do require special attention. For a variable such as narr86, which is zero or one for most men
in the population, a linear model may not be able to adequately capture the functional relationship between
narr86 and the explanatory variables. Moreover, even if a linear model does describe the expected value of
arrests, heteroskedasticity might be a problem. Problems such as these are not mitigated as the sample size
grows, and we will return to them in later chapters.
Key Terms
Asymptotic Bias
Asymptotic Confidence
Interval
Asymptotic Normality
Asymptotic Properties
Asymptotic Standard Error
Asymptotic t Statistics
Asymptotic Variance
Asymptotically Efficient
Auxiliary Regression
Consistency
Inconsistency
Lagrange Multiplier (LM)
Statistic
Large Sample Properties
n-R-Squared Statistic
Score Statistic
Problems
1 In the simple regression model under MLR.1 through MLR.4, we argued that the slope estimator, b^ 1,
is consistent for b1. Using b^ 0 5 y 2 b^ 1x1, show that plim b^ 0 5 b0. [You need to use the consistency of
b^ 1 and the law of large numbers, along with the fact that b0 5 E 1 y 2 2 b1E 1 x1 2 .]
2 Suppose that the model
pctstck 5 b0 1 b1funds 1 b2risktol 1 u
satisfies the first four Gauss-Markov assumptions, where pctstck is the percentage of a worker’s
pension invested in the stock market, funds is the number of mutual funds that the worker can
choose from, and risktol is some measure of risk tolerance (larger risktol means the person has a
higher tolerance for risk). If funds and risktol are positively correlated, what is the inconsistency
| , the slope coefficient in the simple regression of pctstck on funds?
in b
1
58860_ch05_hr_163-180.indd 176
10/18/18 4:25 PM
CHAPTER 5 Multiple Regression Analysis: OLS Asymptotics
177
3 The data set SMOKE contains information on smoking behavior and other variables for a random
sample of single adults from the United States. The variable cigs is the (average) number of cigarettes
smoked per day. Do you think cigs has a normal distribution in the U.S. adult population? Explain.
4 In the simple regression model (5.16), under the first four Gauss-Markov assumptions, we showed that
estimators of the form (5.17) are consistent for the slope, b1. Given such an estimator, define an esti| 5y2b
| x. Show that plim b
| 5b.
mator of b0 by b
0
1
0
0
5 The following histogram was created using the variable score in the data file ECONMATH. Thirty bins
were used to create the histogram, and the height of each cell is the proportion of observations falling
within the corresponding interval. The best-fitting normal distribution—that is, using the sample mean
and sample standard deviation—has been superimposed on the histogram.
.1
proportion in cell
.08
.06
.04
.02
0
20
(i)
(ii)
40
60
80
course score (in percentage form)
100
If you use the normal distribution to estimate the probability that score exceeds 100, would the
answer be zero? Why does your answer contradict the assumption of a normal distribution for
score?
Explain what is happening in the left tail of the histogram. Does the normal distribution fit well
in the left tail?
6 Consider the equation
y 5 b0 1 b1x 1 b2x2 1 u
E1u|x2 5 0
where the explanatory variable x has a standard normal distribution in the population. In particular,
E 1 x 2 5 0, E 1 x2 2 5 Var 1 x 2 5 1, and E 1 x3 2 5 0. This last condition holds because the standard normal
distribution is symmetric about zero. We want to study what we can say about the OLS estimator of b1 we
omit x2 and compute the simple regression estimator of the intercept and slope.
(i) Show that we can write
y 5 a0 1 b1x 1 v
where E(v) 5 0. In particular, find v and the new intercept, a0.
(ii) Show that E(v|x) depends on x unless b2 5 0.
(iii) Show that Cov(x, v) 5 0.
58860_ch05_hr_163-180.indd 177
10/18/18 4:25 PM
178
PART 1 Regression Analysis with Cross-Sectional Data
(iv)
(v)
(vi)
If b^ 1 is the slope coefficient from regression yi on xi, is b^ 1 consistent for b1? Is it unbiased?
­Explain.
Argue that being able to estimate b1 has some value in the following sense: b1 is the partial
­effect of x on y evaluated at x 5 0, the average value of x.
Explain why being able to consistently estimate b1 and b2 is more valuable than just estimating b1.
Computer Exercises
C1 Use the data in WAGE1 for this exercise.
(i) Estimate the equation
wage 5 b0 1 b1educ 1 b2exper 1 b3tenure 1 u.
Save the residuals and plot a histogram.
(ii) Repeat part (i), but with log(wage) as the dependent variable.
(iii) Would you say that Assumption MLR.6 is closer to being satisfied for the level-level model or
the log-level model?
C2 Use the data in GPA2 for this exercise.
(i) Using all 4,137 observations, estimate the equation
colgpa 5 b0 1 b1hsperc 1 b2sat 1 u
and report the results in standard form.
(ii) Reestimate the equation in part (i), using the first 2,070 observations.
(iii) Find the ratio of the standard errors on hsperc from parts (i) and (ii). Compare this with the
­result from (5.10).
C3 In equation (4.42) of Chapter 4, using the data set BWGHT, compute the LM statistic for testing
whether motheduc and fatheduc are jointly significant. In obtaining the residuals for the restricted
model, be sure that the restricted model is estimated using only those observations for which all variables in the unrestricted model are available (see Example 4.9).
C4 Several statistics are commonly used to detect nonnormality in underlying population distributions.
Here we will study one that measures the amount of skewness in a distribution. Recall that any normally distributed random variable is symmetric about its mean; therefore, if we standardize a symmetrically distributed random variable, say z 5 1 y 2 my 2 /sy, where my 5 E 1 y 2 and sy 5 sd 1 y 2 , then
z has mean zero, variance one, and E 1 z3 2 5 0. Given a sample of data 5 yi: i 5 1, . . . , n 6 , we can standardize yi in the sample by using zi 5 1 yi 2 m^ y 2 /s^ y, where m^ y is the sample mean and s^ y is the sample
standard deviation. (We ignore the fact that these are estimates based on the sample.) A sample statistic
n
that measures skewness is n 21 g i51z3i , or where n is replaced with (n 21) as a degrees-of-freedom
­adjustment. If y has a normal distribution in the population, the skewness measure in the sample for the
standardized values should not differ significantly from zero.
(i) First use the data set 401KSUBS, keeping only observations with fsize 5 1. Find the skewness
measure for inc. Do the same for log(inc). Which variable has more skewness and therefore
seems less likely to be normally distributed?
(ii) Next use BWGHT2. Find the skewness measures for bwght and log(bwght). What do you
­conclude?
(iii) Evaluate the following statement: “The logarithmic transformation always makes a positive
variable look more normally distributed.”
(iv) If we are interested in the normality assumption in the context of regression, should we be
evaluating the unconditional distributions of y and log(y)? Explain.
58860_ch05_hr_163-180.indd 178
10/18/18 4:25 PM
CHAPTER 5 Multiple Regression Analysis: OLS Asymptotics
179
C5 Consider the analysis in Computer Exercise C11 in Chapter 4 using the data in HTV, where educ is the
dependent variable in a regression.
(i) How many different values are taken on by educ in the sample? Does educ have a continuous
distribution?
(ii) Plot a histogram of educ with a normal distribution overlay. Does the distribution of educ
­appear anything close to normal?
(iii) Which of the CLM assumptions seems clearly violated in the model
educ 5 b0 1 b1motheduc 1 b2fatheduc 1 b3abil 1 b4abil2 1 u?
How does this violation change the statistical inference procedures carried out in Computer
­Exercise C11 in Chapter 4?
C6 Use the data in ECONMATH to answer this question.
(i) Logically, what are the smallest and largest values that can be taken on by the variable score?
What are the smallest and largest values in the sample?
(ii) Consider the linear model
score 5 b0 1 b1colgpa 1 b2actmth 1 b3acteng 1 u.
Why cannot Assumption MLR.6 hold for the error term u? What consequences does this have
for using the usual t statistic to test H0: b3 5 0?
(iii) Estimate the model from part (ii) and obtain the t statistic and associated p-value for testing
H0: b3 5 0. How would you defend your findings to someone who makes the following statement: “You cannot trust that p-value because clearly the error term in the equation cannot have
a normal distribution.”
Appendix 5A
Asymptotic Normality of OLS
We sketch a proof of the asymptotic normality of OLS [Theorem 5.2(i)] in the simple regression
case. Write the simple regression model as in equation (5.16). Then, by the usual algebra of simple
regression, we can write
!n 1 b^ 1 2 b1 2 5 1 1/s2x 2 c n21/2 a 1 xi 2 x 2 ui d ,
n
i51
where we use to denote the sample variance of 5 xi: i 5 1, 2, . . . , n 6 . By the law of large nump
bers (see Math Refresher C), s2x S s2x 5 Var 1 x 2 . Assumption MLR.3 rules out perfect collinearity,
which means that Var(x) > 0 (xi varies in the sample, and therefore x is not constant in the populan
n
n
tion). Next, n21/2 g i51 1 xi 2 x 2 ui 5 n21/2 g i51 1 xi 2 m 2 ui 1 1 m 2 x 2 [n 21/2 g i51ui], where m 5 E 1 x 2
is the population mean of x. Now 5 ui 6 is a sequence of i.i.d. random variables with mean zero and
n
variance s2, and so n21/2 g i51ui converges to the Normal(0,s2) distribution as n S `; this is just
the central limit theorem from Math Refresher C. By the law of large numbers, plim 1 u 2 x 2 5 0.
A standard result in asymptotic theory is that if plim 1 wn 2 5 0 and zn has an asymptotic normal
distribution, then plim 1 wnzn 2 5 0. [See Wooldridge (2010, Chapter 3) for more discussion.] This
n
implies that 1 m 2 x 2 [n21/2 g i51ui ] has zero plim. Next, { 1 xi 2 m 2 ui: i 5 1, 2, p} is an indefinite sequence of i.i.d. random variables with mean zero—because u and x are uncorrelated under
Assumption MLR.4—and variance s2s2x by the homoskedasticity Assumption MLR.5. Therefore,
s2x
58860_ch05_hr_163-180.indd 179
10/18/18 4:25 PM
180
PART 1 Regression Analysis with Cross-Sectional Data
n
n21/2 g i51 1 xi 2 m 2 ui has an asymptotic Normal 1 0,s2s2x 2 distribution. We just showed that the
n
n
­difference between n21/2 g i51 1 xi 2 x 2 ui and n21/2 g i51 1 xi 2 m 2 ui has zero plim. A result in asymptotic theory is that if zn has an asymptotic normal distribution and plim 1 vn 2 zn 2 5 0, then vn has the
n
same asymptotic normal distribution as zn. It follows that n21/2 g i51 1 xi 2 x 2 ui also has an asymptotic
2 2
Normal 1 0,s sx 2 distribution. Putting all of the pieces together gives
"n 1 b^ 1 2 b1 2 5 1 1/s2x 2 c n21/2 a 1 xi 2 x 2 ui d
n
i51
1 [ 1 1/s2x 2 2 1 1/s2x )] c n21/2 a 1 xi 2 x 2 ui d ,
n
i51
and because plim 1 1/s2x 2 5 1/s2x , the second term has zero plim. Therefore, the asymptotic distribution of !n 1 b^ 1 2 b1 2 is Normal 1 0, 5 s2s2x 6 / 5 s2x 6 2 2 5 Normal 1 0,s2/s2x 2 . This completes the proof
in the simple regression case, as a21 5 s2x in this case. See Wooldridge (2010, Chapter 4) for the
general case.
58860_ch05_hr_163-180.indd 180
10/18/18 4:25 PM
chapter
6
Multiple Regression
Analysis: Further Issues
T
his chapter brings together several issues in multiple regression analysis that we could
not conveniently cover in earlier chapters. These topics are not as fundamental as the material
in Chapters 3 and 4, but they are important for applying multiple regression to a broad range
of empirical problems.
6-1 Effects of Data Scaling on OLS Statistics
In Chapter 2 on bivariate regression, we briefly discussed the effects of changing the units of measurement on the OLS intercept and slope estimates. We also showed that changing the units of measurement did not affect R-squared. We now return to the issue of data scaling and examine the effects
of rescaling the dependent or independent variables on standard errors, t statistics, F statistics, and
confidence intervals.
We will discover that everything we expect to happen, does happen. When variables are rescaled,
the coefficients, standard errors, confidence intervals, t statistics, and F statistics change in ways that
preserve all measured effects and testing outcomes. Although this is no great surprise—in fact, we
would be very worried if it were not the case—it is useful to see what occurs explicitly. Often, data
scaling is used for cosmetic purposes, such as to reduce the number of zeros after a decimal point in
an estimated coefficient. By judiciously choosing units of measurement, we can improve the appearance of an estimated equation while changing nothing that is essential.
We could treat this problem in a general way, but it is much better illustrated with examples.
Likewise, there is little value here in introducing an abstract notation.
181
58860_ch06_hr_181-219.indd 181
10/18/18 4:26 PM
182
PART 1 Regression Analysis with Cross-Sectional Data
We begin with an equation relating infant birth weight to cigarette smoking and family income:
bwght 5 b^ 0 1 b^ 1cigs 1 b^ 2 faminc,
[6.1]
where
bwght 5 child birth weight, in ounces.
cigs 5 number of cigarettes smoked by the mother while pregnant, per day.
faminc 5 annual family income, in thousands of dollars.
The estimates of this equation, obtained using the data in BWGHT, are given in the first column of
Table 6.1. Standard errors are listed in parentheses. The estimate on cigs says that if a woman smoked
five more cigarettes per day, birth weight is predicted to be about .4634 1 5 2 5 2.317 ounces less. The
t statistic on cigs is −5.06, so the variable is very statistically significant.
Now, suppose that we decide to measure birth weight in pounds, rather than in ounces. Let
bwghtlbs 5 bwght/16 be birth weight in pounds. What happens to our OLS statistics if we use this
as the dependent variable in our equation? It is easy to find the effect on the coefficient estimates by
simple manipulation of equation (6.1). Divide this entire equation by 16:
bwght/16 5 b^ 0 /16 1 1 b^ 1/16 2 cigs 1 1 b^ 2 /16 2 faminc.
Because the left-hand side is birth weight in pounds, it follows that each new coefficient will be the
corresponding old coefficient divided by 16. To verify this, the regression of bwghtlbs on cigs, and
faminc is reported in column (2) of Table 6.1. Up to the reported digits (and any digits beyond),
the intercept and slopes in column (2) are just those in column (1) divided by 16. For example, the
­coefficient on cigs is now −.0289; this means that if cigs were higher by five, birth weight would be
.0289 1 5 2 5 .1445 pounds lower. In terms of ounces, we have .1445 1 16 2 5 2.312, which is slightly
different from the 2.317 we obtained earlier due to rounding error. The point is, after the effects are
transformed into the same units, we get exactly the same answer, regardless of how the dependent
variable is measured.
What about statistical significance? As we expect, changing the dependent variable from ounces
to pounds has no effect on how statistically important the independent variables are. The standard
errors in column (2) are 16 times smaller than those in column (1). A few quick calculations show
Table 6.1 Effects of Data Scaling
Dependent Variable
(1) bwght
(2) bwghtlbs
(3) bwght
−.4634
(.0916)
−.0289
(.0057)
—
packs
—
—
−9.268
(1.832)
faminc
.0927
(.0292)
.0058
(.0018)
Independent Variables
cigs
intercept
Observations
R-Squared
SSR
SER
58860_ch06_hr_181-219.indd 182
116.974
(1.049)
1,388
.0298
557,485.51
20.063
7.3109
(.0656)
1,388
.0298
2,177.6778
1.2539
.0927
(.0292)
116.974
(1.049)
1,388
.0298
557,485.51
20.063
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
183
that the t statistics in column (2) are indeed identical to the t statistics in column (1). The endpoints
for the confidence intervals in column (2) are just the endpoints in column (1) divided by 16. This
is because the CIs change by the same factor as the standard errors. [Remember that the 95% CI here
is b^ j 6 1.96 se 1 b^ j 2 .]
In terms of goodness-of-fit, the R-squareds from the two regressions are identical, as should be
the case. Notice that the sum of squared residuals (SSR) and the standard error of the regression
(SER) do differ across equations. These differences are easily explained. Let u^ i denote the residual for
observation i in the original equation (6.1). Then the residual when bwghtlbs is the dependent variable
is simply u^ i/16. Thus, the squared residual in the second equation is 1 u^ i/16 2 2 5 u^ 2i /256. This is why
the SSR in column (2) is equal to the SSR in column (1) divided by 256.
Because SER 5 s^ 5 "SSR/ 1 n 2 k 2 1 2 5 "SSR/1,385, the SER in column (2) is 16 times
smaller than that in column (1). Another way to think about this is that the error in the equation with
bwghtlbs as the dependent variable has a standard deviation 16 times smaller than the standard deviation of the original error. This does not mean that we have reduced the error by changing how birth
weight is measured; the smaller SER simply reflects a difference in units of measurement.
Next, let us return the dependent variable to its original units: bwght is measured in ounces.
Instead, let us change the unit of measurement of one of the independent variables, cigs. Define packs
to be the number of packs of cigarettes smoked per day. Thus, packs 5 cigs/20. What happens to the
coefficients and other OLS statistics now? Well, we can write
bwght 5 b^ 0 1 1 20b^ 1 2 1 cigs/20 2 1 b^ 2 faminc 5 b^ 0 1 1 20b^ 1 2 packs 1 b^ 2 faminc.
Thus, the intercept and slope coefficient on faminc are unchanged, but the coefficient on packs is
20 times that on cigs. This is intuitively appealing. The results from the regression of bwght on
packs and faminc are in column (3) of Table 6.1.
Incidentally, remember that it would make no sense
G o i n g F u rt h e r 6 . 1
to include both cigs and packs in the same equation; this would induce perfect multicollinearity and
In the original birth weight equation (6.1),
would have no interesting meaning.
suppose that faminc is measured in dollars
rather than in thousands of dollars. Thus,
Other than the coefficient on packs, there is one
define the variable fincdol 5 1,000 # faminc.
other statistic in column (3) that differs from that in
How will the OLS statistics change when
column (1): the standard error on packs is 20 times
fincdol is substituted for faminc? For the
larger than that on cigs in column (1). This means
purpose of presenting the regression rethat the t statistic for testing the significance of
sults, do you think it is better to measure
cigarette smoking is the same whether we measure
income in dollars or in thousands of dollars?
smoking in terms of cigarettes or packs. This is only
natural.
The previous example spells out most of the possibilities that arise when the dependent and independent variables are rescaled. Rescaling is often done with dollar amounts in economics, especially
when the dollar amounts are very large.
In Chapter 2, we argued that, if the dependent variable appears in logarithmic form, changing the
unit of measurement does not affect the slope coefficient. The same is true here: changing the unit
of measurement of the dependent variable, when it appears in logarithmic form, does not affect any
of the slope estimates. This follows from the simple fact that log 1 c1yi 2 5 log 1 c1 2 1 log 1 yi 2 for any
constant c1 . 0. The new intercept will be log 1 c1 2 1 b^ 0. Similarly, changing the unit of measurement
of any xj, where log 1 xj 2 appears in the regression, only affects the intercept. This corresponds to what
we know about percentage changes and, in particular, elasticities: they are invariant to the units of
measurement of either y or the xj. For example, if we had specified the dependent variable in (6.1) to
be log(bwght), estimated the equation, and then reestimated it with log(bwghtlbs) as the dependent
variable, the coefficients on cigs and faminc would be the same in both regressions; only the intercept
would be different.
58860_ch06_hr_181-219.indd 183
10/18/18 4:26 PM
184
PART 1 Regression Analysis with Cross-Sectional Data
6-1a Beta Coefficients
Sometimes, in econometric applications, a key variable is measured on a scale that is difficult to interpret. Labor economists often include test scores in wage equations, and the scale on which these tests
are scored is often arbitrary and not easy to interpret (at least for economists!). In almost all cases,
we are interested in how a particular individual’s score compares with the population. Thus, instead
of asking about the effect on hourly wage if, say, a test score is 10 points higher, it makes more sense
to ask what happens when the test score is one standard deviation higher.
Nothing prevents us from seeing what happens to the dependent variable when an independent
variable in an estimated model increases by a certain number of standard deviations, assuming that
we have obtained the sample standard deviation of the independent variable (which is easy in most
regression packages). This is often a good idea. So, for example, when we look at the effect of a standardized test score, such as the SAT score, on college GPA, we can find the standard deviation of SAT
and see what happens when the SAT score increases by one or two standard deviations.
Sometimes, it is useful to obtain regression results when all variables involved, the dependent as
well as all the independent variables, have been standardized. A variable is standardized in the sample by
subtracting off its mean and dividing by its standard deviation (see Math Refresher C). This means that
we compute the z-score for every variable in the sample. Then, we run a regression using the z-scores.
Why is standardization useful? It is easiest to start with the original OLS equation, with the variables in their original forms:
yi 5 b^ 0 1 b^ 1xi1 1 b^ 2xi2 1 p 1 b^ kxik 1 u^ i.
[6.2]
We have included the observation subscript i to emphasize that our standardization is applied to all
sample values. Now, if we average (6.2), use the fact that the u^ i have a zero sample average, and subtract the result from (6.2), we get
yi 2 y 5 b^ 1 1 xi1 2 x1 2 1 b^ 2 1 xi2 2 x2 2 1 p 1 b^ k 1 xik 2 xk 2 1 u^ i.
Now, let s^ y be the sample standard deviation for the dependent variable, let s^ 1 be the sample sd for x1,
let s^ 2 be the sample sd for x2, and so on. Then, simple algebra gives the equation
1 yi 2 y 2 /s^ y 5 1 s^ 1/s^ y 2 b^ 1 3 1 xi1 2 x1 2 /s^ 1 4 1 p
1 1 s^ k/s^ y 2 b^ k 3 1 xik 2 xk 2 /s^ k 4 1 1 u^ i/s^ y 2 .
[6.3]
zy 5 b^ 1z1 1 b^ 2z2 1 p 1 b^ kzk 1 error,
[6.4]
Each variable in (6.3) has been standardized by replacing it with its z-score, and this has resulted in
new slope coefficients. For example, the slope coefficient on 1 xi1 2 x1 2 /s^ 1 is 1 s^ 1 /s^ y 2 b^ 1. This is simply the original coefficient, b^ 1, multiplied by the ratio of the standard deviation of x1 to the standard
deviation of y. The intercept has dropped out altogether.
It is useful to rewrite (6.3), dropping the i subscript, as
where zy denotes the z-score of y, z1 is the z-score of x1, and so on. The new coefficients are
b^ j 5 1 s^ j/s^ y 2 b^ j for j 5 1, . . . , k.
[6.5]
These b^ j are traditionally called standardized coefficients or beta coefficients. (The latter name is
more common, which is unfortunate because we have been using beta hat to denote the usual OLS
estimates.)
58860_ch06_hr_181-219.indd 184
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
185
Beta coefficients receive their interesting meaning from equation (6.4): if x1 increases by one
standard deviation, then y^ changes by b^ 1 standard deviations. Thus, we are measuring effects not in
terms of the original units of y or the xj, but in standard deviation units. Because it makes the scale of
the regressors irrelevant, this equation puts the explanatory variables on equal footing. In a standard
OLS equation, it is not possible to simply look at the size of different coefficients and conclude that
the explanatory variable with the largest coefficient is “the most important.” We just saw that the
magnitudes of coefficients can be changed at will by changing the units of measurement of the xj.
But, when each xj has been standardized, comparing the magnitudes of the resulting beta coefficients
is more compelling. When the regression equation has only a single explanatory variable, x1, its standardized coefficient is simply the sample correlation coefficient between y and x1, which means it
must lie in the range −1 to 1.
Even in situations in which the coefficients are easily interpretable—say, the dependent variable
and independent variables of interest are in logarithmic form, so the OLS coefficients of interest are
estimated elasticities—there is still room for computing beta coefficients. Although elasticities are
free of units of measurement, a change in a particular explanatory variable by, say, 10% may represent a larger or smaller change over a variable’s range than changing another explanatory variable by
10%. For example, in a state with wide income variation but relatively little variation in spending per
student, it might not make much sense to compare performance elasticities with respect to the income
and spending. Comparing beta coefficient magnitudes can be helpful.
To obtain the beta coefficients, we can always standardize y, x1, . . . , xk and then run the OLS
regression of the z-score of y on the z-scores of x1, . . . , xk—where it is not necessary to include an
intercept, as it will be zero. This can be tedious with many independent variables. Many regression
packages provide beta coefficients via a simple command. The following example illustrates the use
of beta coefficients.
Example 6.1
Effects of Pollution on Housing Prices
We use the data from Example 4.5 (in the file HPRICE2) to illustrate the use of beta coefficients.
Recall that the key independent variable is nox, a measure of the nitrogen oxide in the air over each
community. One way to understand the size of the pollution effect—without getting into the science
underlying nitrogen oxide’s effect on air quality—is to compute beta coefficients. (An alternative
­approach is contained in Example 4.5: we obtained a price elasticity with respect to nox by using
price and nox in logarithmic form.)
The population equation is the level-level model
price 5 b0 1 b1nox 1 b2crime 1 b3rooms 1 b4dist 1 b5stratio 1 u,
where all the variables except crime were defined in Example 4.5; crime is the number of reported
crimes per capita. The beta coefficients are reported in the following equation (so each variable has
been converted to its z-score):
zprice 5 2.340 znox 2 .143 zcrime 1 .514 zrooms 2 .235 zdist 2 .270 zstratio.
This equation shows that a one standard deviation increase in nox decreases price by .34 standard
­deviation; a one standard deviation increase in crime reduces price by .14 standard deviation. Thus,
the same relative movement of pollution in the population has a larger effect on housing prices than
crime does. Size of the house, as measured by number of rooms (rooms), has the largest standardized effect. If we want to know the effects of each independent variable on the dollar value of median
house price, we should use the unstandardized variables.
Whether we use standardized or unstandardized variables does not affect statistical significance:
the t statistics are the same in both cases.
58860_ch06_hr_181-219.indd 185
10/18/18 4:26 PM
186
PART 1 Regression Analysis with Cross-Sectional Data
6-2 More on Functional Form
In several previous examples, we have encountered the most popular device in econometrics for
­allowing nonlinear relationships between the explained and explanatory variables: using logarithms
for the dependent or independent variables. We have also seen models containing quadratics in some
explanatory variables, but we have yet to provide a systematic treatment of them. In this section, we
cover some variations and extensions on functional forms that often arise in applied work.
6-2a More on Using Logarithmic Functional Forms
We begin by reviewing how to interpret the parameters in the model
log 1 price 2 5 b0 1 b1log 1 nox 2 1 b2rooms 1 u,
[6.6]
where these variables are taken from Example 4.5. Recall that throughout the text log(x) is the natural
log of x. The coefficient b1 is the elasticity of price with respect to nox (pollution). The coefficient b2
is the change in log( price), when Drooms 5 1; as we have seen many times, when multiplied by 100,
this is the approximate percentage change in price. Recall that 100 # b2 is sometimes called the semielasticity of price with respect to rooms.
When estimated using the data in HPRICE2, we obtain
log 1 price 2 5 9.23 2 .718 log 1 nox 2 1 .306 rooms
1 0.19 2 1 .066 2
1 .019 2
2
n 5 506, R 5 .514.
[6.7]
Thus, when nox increases by 1%, price falls by .718%, holding only rooms fixed. When rooms
­increases by one, price increases by approximately 100 1 .306 2 5 30.6%.
The estimate that one more room increases price by about 30.6% turns out to be somewhat
­inaccurate for this application. The approximation error occurs because, as the change in log(y)
­becomes larger and larger, the approximation %Dy < 100 # Dlog 1 y 2 becomes more and more inaccurate. Fortunately, a simple calculation is available to compute the exact percentage change.
To describe the procedure, we consider the general estimated model
log 1 y 2 5 b^ 0 1 b^ 1log 1 x1 2 1 b^ 2x2.
(Adding additional independent variables does not change the procedure.) Now, fixing x1, we have
Dlog 1 y 2 5 b^ 2Dx2. Using simple algebraic properties of the exponential and logarithmic functions
gives the exact percentage change in the predicted y as
%Dy^ 5 100 # 3 exp 1 b^ 2Dx2 2 2 1 4 ,
[6.8]
%Dy^ 5 100 # 3 exp 1 b^ 2 2 2 1 4 .
[6.9]
where the multiplication by 100 turns the proportionate change into a percentage change. When
Dx2 5 1,
Applied to the housing price example with x2 5 rooms and b^ 2 5 .306, %Dprice 5 100 3 exp 1 .306 2
2 1 4 5 35.8%, which is notably larger than the approximate percentage change, 30.6%, obtained
directly from (6.7). {Incidentally, this is not an unbiased estimator because exp(·) is a nonlinear
function; it is, however, a consistent estimator of 100 3 exp 1 b2 2 2 1 4 . This is because the probability
limit passes through continuous functions, while the expected value operator does not. See Math
Refresher C.}
58860_ch06_hr_181-219.indd 186
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
187
The adjustment in equation (6.8) is not as crucial for small percentage changes. For example,
when we include the student-teacher ratio in equation (6.7), its estimated coefficient is −.052, which
means that if stratio increases by one, price decreases by approximately 5.2%. The exact proportionate change is exp 1 2.052 2 21 < 2.051, or 25.1%. On the other hand, if we increase stratio by five,
then the approximate percentage change in price is −26%, while the exact change obtained from
equation (6.8) is 100 3 exp 1 2.26 2 21 4 < 222.9%.
The logarithmic approximation to percentage changes has an advantage that justifies its reporting
even when the percentage change is large. To describe this advantage, consider again the effect on
price of changing the number of rooms by one. The logarithmic approximation is just the coefficient
on rooms in equation (6.7) multiplied by 100, namely, 30.6%. We also computed an estimate of the
exact percentage change for increasing the number of rooms by one as 35.8%. But what if we want to
estimate the percentage change for decreasing the number of rooms by one? In equation (6.8) we take
D x2 5 21 and b^ 2 5 .306, and so %Dprice 5 100 3 exp 1 2.306 2 21 4 5 226.4, or a drop of 26.4%.
Notice that the approximation based on using the coefficient on rooms is between 26.4 and 35.8—an
outcome that always occurs. In other words, simply using the coefficient (multiplied by 100) gives us
an estimate that is always between the absolute value of the estimates for an increase and a decrease.
If we are specifically interested in an increase or a decrease, we can use the calculation based on
equation (6.8).
The point just made about computing percentage changes is essentially the one made in introductory economics when it comes to computing, say, price elasticities of demand based on large price
changes: the result depends on whether we use the beginning or ending price and quantity in computing the percentage changes. Using the logarithmic approximation is similar in spirit to calculating an
arc elasticity of demand, where the averages of prices and quantities are used in the denominators in
computing the percentage changes.
We have seen that using natural logs leads to coefficients with appealing interpretations, and we
can be ignorant about the units of measurement of variables appearing in logarithmic form because
the slope coefficients are invariant to rescalings. There are several other reasons logs are used so much
in applied work. First, when y . 0, models using log(y) as the dependent variable often satisfy the
CLM assumptions more closely than models using the level of y. Strictly positive variables often have
conditional distributions that are heteroskedastic or skewed; taking the log can mitigate, if not eliminate, both problems.
Another potential benefit of using logs is that taking the log of a variable often narrows its range.
This is particularly true of variables that can be large monetary values, such as firms’ annual sales or
baseball players’ salaries. Population variables also tend to vary widely. Narrowing the range of the
dependent and independent variables can make OLS estimates less sensitive to outlying (or extreme)
values; we take up the issue of outlying observations in Chapter 9.
However, one must not indiscriminately use the logarithmic transformation because in some cases
it can actually create extreme values. An example is when a variable y is between zero and one (such
as a proportion) and takes on values close to zero. In this case, log(y) (which is necessarily negative)
can be very large in magnitude whereas the original variable, y, is bounded between zero and one.
There are some standard rules of thumb for taking logs, although none is written in stone. When
a variable is a positive dollar amount, the log is often taken. We have seen this for variables such
as wages, salaries, firm sales, and firm market value. Variables such as population, total number of
employees, and school enrollment often appear in logarithmic form; these have the common feature
of being large integer values.
Variables that are measured in years—such as education, experience, tenure, age, and so on—
usually appear in their original form. A variable that is a proportion or a percent—such as the
unemployment rate, the participation rate in a pension plan, the percentage of students passing a
standardized exam, and the arrest rate on reported crimes—can appear in either original or logarithmic form, although there is a tendency to use them in level forms. This is because any regression
58860_ch06_hr_181-219.indd 187
10/18/18 4:26 PM
188
PART 1 Regression Analysis with Cross-Sectional Data
coefficients involving the original variable—whether it is the dependent or independent variable—
will have a percentage point change interpretation. (See Math Refresher A for a review of the distinction between a percentage change and a percentage point change.) If we use, say, log(unem) in a
regression, where unem is the percentage of unemployed individuals, we must be very careful to distinguish between a percentage point change and a percentage change. Remember, if unem goes from
8 to 9, this is an increase of one percentage point, but a 12.5% increase from the initial unemployment level. Using the log means that we are looking at the percentage change in the unemployment
rate: log 1 9 2 2 log 1 8 2 < .118 or 11.8%, which is the logarithmic approximation to the actual 12.5%
increase.
One limitation of the log is that it cannot be used if a variable takes on zero or negative values. In
cases where a variable y is nonnegative but can take on the value 0, log(11y) is sometimes used. The
percentage change interpretations are often closely
preserved, except for changes beginning at y 5 0
G o i n g F u rt h e r 6 . 2
(where the percentage change is not even defined).
Generally, using log(11y) and then interpreting the
Suppose that the annual number of drunk
estimates as if the variable were log(y) is acceptdriving arrests is determined by
able when the data on y contain relatively few zeros.
log 1 arrests 2 5 b0 1 b1log 1 pop 2
An example might be where y is hours of training
1 b2age16_25 1 other
per employee for the population of manufacturing
factors,
firms, if a large fraction of firms provides training
where age 16_25 is the proportion of the
to at least one worker. Technically, however, log (11y)
population between 16 and 25 years of
cannot be normally distributed (although it might
age. Show that b2 has the following (ceteris
be less heteroskedastic than y). Useful, albeit more
paribus) interpretation: it is the percentage
advanced, alternatives are the Tobit and Poisson
change in arrests when the percentage of
models in Chapter 17.
the people aged 16 to 25 increases by one
One drawback to using a dependent variable in
percentage point.
logarithmic form is that it is more difficult to predict
the original variable. The original model allows us
to predict log(y), not y. Nevertheless, it is fairly easy
to turn a prediction for log(y) into a prediction for y (see Section 6-4). A related point is that it is not
legitimate to compare R-squareds from models where y is the dependent variable in one case and
log(y) is the dependent variable in the other. These measures explain variations in different variables.
We discuss how to compute comparable goodness-of-fit measures in Section 6-4.
6-2b Models with Quadratics
Quadratic functions are also used quite often in applied economics to capture decreasing or increasing marginal effects. You may want to review properties of quadratic functions in Math Refresher A.
In the simplest case, y depends on a single observed factor x, but it does so in a quadratic fashion:
y 5 b0 1 b1x 1 b2x2 1 u.
For example, take y 5 wage and x 5 exper. As we discussed in Chapter 3, this model falls outside of
simple regression analysis but is easily handled with multiple regression.
It is important to remember that b1 does not measure the change in y with respect to x; it makes
no sense to hold x2 fixed while changing x. If we write the estimated equation as
y^ 5 b^ 0 1 b^ 1x 1 b^ 2x2,
[6.10]
Dy^ < 1 b^ 1 1 2b^ 2x 2 Dx, so Dy^ /Dx < b^ 1 1 2b^ 2x.
[6.11]
then we have the approximation
58860_ch06_hr_181-219.indd 188
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
189
This says that the slope of the relationship between x and y depends on the value of x; the estimated
slope is b^ 1 1 2b^ 2x. If we plug in x 5 0, we see that b^ 1 can be interpreted as the approximate slope in
going from x 5 0 to x 5 1. After that, the second term, 2b^ 2x, must be accounted for.
If we are only interested in computing the predicted change in y given a starting value for x and a
change in x, we could use (6.10) directly: there is no reason to use the calculus approximation at all.
However, we are usually more interested in quickly summarizing the effect of x on y, and the interpretation of b^ 1 and b^ 2 in equation (6.11) provides that summary. Typically, we might plug in the average
value of x in the sample, or some other interesting values, such as the median or the lower and upper
quartile values.
In many applications, b^ 1 is positive and b^ 2 is negative. For example, using the wage data in
WAGE1, we obtain
wage 5 3.73 1 .298 exper 2 .0061 exper2
1 .35 2 1 .041 2
1 .0009 2
[6.12]
2
n 5 526, R 5 .093.
This estimated equation implies that exper has a diminishing effect on wage. The first year of experience is worth roughly 30¢ per hour ($.298). The second year of experience is worth less [about
.298 2 2 1 .0061 2 1 1 2 < .286, or 28.6¢, according to the approximation in (6.11) with x 5 1]. In going
from 10 to 11 years of experience, wage is predicted to increase by about .298 2 2 1 .0061 2 1 10 2 5 .176,
or 17.6¢. And so on.
When the coefficient on x is positive and the coefficient on x2 is negative, the quadratic has a
parabolic shape. There is always a positive value of x where the effect of x on y is zero; before this
point, x has a positive effect on y; after this point, x has a negative effect on y. In practice, it can be
important to know where this turning point is.
In the estimated equation (6.10) with b^ 1 . 0 and b^ 2 , 0, the turning point (or maximum of
the function) is always achieved at the coefficient on x over twice the absolute value of the coefficient
on x2:
xp 5 0 b^ 1/ 1 2b^ 2 2 0 .
[6.13]
In the wage example, xp 5 experp is .298/ 3 2 1 .0061 2 4 < 24.4. (Note how we just drop the minus sign
on −.0061 in doing this calculation.) This quadratic relationship is illustrated in Figure 6.1.
In the wage equation (6.12), the return to experience becomes zero at about 24.4 years. What
should we make of this? There are at least three possible explanations. First, it may be that few
people in the sample have more than 24 years of experience, and so the part of the curve to the
right of 24 can be ignored. The cost of using a quadratic to capture diminishing effects is that
the quadratic must eventually turn around. If this point is beyond all but a small percentage of
the people in the sample, then this is not of much concern. But in the data set WAGE1, about 28%
of the people in the sample have more than 24 years of experience; this is too high a percentage
to ignore.
It is possible that the return to exper really becomes negative at some point, but it is hard to
believe that this happens at 24 years of experience. A more likely possibility is that the estimated
effect of exper on wage is biased because we have controlled for no other factors, or because the
functional relationship between wage and exper in equation (6.12) is not entirely correct. Computer
Exercise C2 asks you to explore this possibility by controlling for education, in addition to using
log(wage) as the dependent variable.
When a model has a dependent variable in logarithmic form and an explanatory variable entering
as a quadratic, some care is needed in reporting the partial effects. The following example also shows
that the quadratic can have a U-shape, rather than a parabolic shape. A U-shape arises in equation
(6.10) when b^ 1 is negative and b^ 2 is positive; this captures an increasing effect of x on y.
58860_ch06_hr_181-219.indd 189
10/18/18 4:26 PM
190
PART 1 Regression Analysis with Cross-Sectional Data
Figure 6.1 Quadratic relationship between wage and exper.
wage
7.37
3.73
24.4
Example 6.2
exper
Effects of Pollution on Housing Prices
We modify the housing price model from Example 4.5 to include a quadratic term in rooms:
log 1 price 2 5 b0 1 b1log 1 nox 2 1 b2log 1 dist 2 1 b3rooms
1 b4rooms2 1 b5stratio 1 u.
[6.14]
The model estimated using the data in HPRICE2 is
log 1 price 2 5 13.39 2 .902 log 1 nox 2 2 .087 log 1 dist 2
1 .57 2 1 .115 2
1 .043 2
2 .545 rooms 1 .062 rooms2 2 .048 stratio
1 .165 2
1 .013 2
1 .006 2
2
n 5 506, R 5 .603.
The quadratic term rooms2 has a t statistic of about 4.77, and so it is very statistically significant. But
what about interpreting the effect of rooms on log(price)? Initially, the effect appears to be strange.
Because the coefficient on rooms is negative and the coefficient on rooms2 is positive, this equation
literally implies that, at low values of rooms, an additional room has a negative effect on log(price).
At some point, the effect becomes positive, and the quadratic shape means that the semi-elasticity of
price with respect to rooms is increasing as rooms increases. This situation is shown in Figure 6.2.
We obtain the turnaround value of rooms using equation (6.13) (even though b^ 1 is negative and
^b2 is positive). The absolute value of the coefficient on rooms, .545, divided by twice the coefficient
on rooms2, .062, gives roomsp 5 .545/ 3 2 1 .062 2 4 < 4.4; this point is labeled in Figure 6.2.
Do we really believe that starting at three rooms and increasing to four rooms actually reduces a
house’s expected value? Probably not. It turns out that only five of the 506 communities in the sample
58860_ch06_hr_181-219.indd 190
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
191
Figure 6.2 log(price) as a quadratic function of rooms.
log(price)
4.4
rooms
have houses averaging 4.4 rooms or less, about 1% of the sample. This is so small that the quadratic
to the left of 4.4 can, for practical purposes, be ignored. To the right of 4.4, we see that adding another
room has an increasing effect on the percentage change in price:
and so
Dlog 1 price 2 < 5 3 2.545 1 2 1 .062 2 4 rooms 6 Drooms
%Dprice < 100 5 3 2.545 1 2 1 .062 2 4 rooms 6 Drooms
5 1 254.5 1 12.4 rooms 2 Drooms.
Thus, an increase in rooms from, say, five to six increases price by about 254.5 1 12.4 1 5 2 5 7.5%;
the increase from six to seven increases price by roughly 254.5 1 12.4 1 6 2 5 19.9%. This is a very
strong increasing effect.
The strong increasing effect of rooms on log(price) in this example illustrates an important lesson: one cannot simply look at the coefficient on the quadratic term—in this case, .062—and declare
that it is too small to bother with, based only on its magnitude. In many applications with quadratics,
the coefficient on the squared variable has one or more zeros after the decimal point: after all, this
coefficient measures how the slope is changing as x (rooms) changes. A seemingly small coefficient
can have practically important consequences, as we just saw. As a general rule, one must compute the
partial effect and see how it varies with x to determine if the quadratic term is practically important. In
doing so, it is useful to compare the changing slope implied by the quadratic model with the constant
slope obtained from the model with only a linear term. If we drop rooms2 from the equation, the coefficient on rooms becomes about .255, which implies that each additional room—starting from any
number of rooms—increases median price by about 25.5%. This is very different from the quadratic
model, where the effect becomes 25.5% at rooms 5 6.45 but changes rapidly as rooms gets smaller
or larger. For example, at rooms 5 7, the return to the next room is about 32.3%.
58860_ch06_hr_181-219.indd 191
10/18/18 4:26 PM
192
PART 1 Regression Analysis with Cross-Sectional Data
What happens generally if the coefficients on the level and squared terms have the same sign
(either both positive or both negative) and the explanatory variable is necessarily nonnegative (as in
the case of rooms or exper)? In either case, there is no turning point for values x . 0. For example,
if b1 and b2 are both positive, the smallest expected value of y is at x 5 0, and increases in x always
have a positive and increasing effect on y. (This is also true if b1 5 0 and b2 . 0, which means that
the partial effect is zero at x 5 0 and increasing as x increases.) Similarly, if b1 and b2 are both negative, the largest expected value of y is at x 5 0, and increases in x have a negative effect on y, with the
magnitude of the effect increasing as x gets larger.
The general formula for the turning point of any quadratic is xp 5 2b^ 1/ 1 2b^ 2 2 , which leads to a
positive value if b^ 1 and b^ 2 have opposite signs and a negative value when b^ 1 and b^ 2 have the same
sign. Knowing this simple formula is useful in cases where x may take on both positive and negative
values; one can compute the turning point and see if it makes sense, taking into account the range of
x in the sample.
There are many other possibilities for using quadratics along with logarithms. For example, an
extension of (6.14) that allows a nonconstant elasticity between price and nox is
log 1 price 2 5 b0 1 b1log 1 nox 2 1 b2 3 log 1 nox 2 4 2
1 b3crime 1 b4rooms 1 b5rooms2 1 b6stratio 1 u.
[6.15]
If b2 5 0, then b1 is the elasticity of price with respect to nox. Otherwise, this elasticity depends on
the level of nox. To see this, we can combine the arguments for the partial effects in the quadratic and
logarithmic models to show that
%Dprice < 3 b1 1 2b2log 1 nox 2 4 %Dnox;
[6.16]
therefore, the elasticity of price with respect to nox is b1 1 2b2log 1 nox 2 , so that it depends on
log(nox).
Finally, other polynomial terms can be included in regression models. Certainly, the quadratic is
seen most often, but a cubic and even a quartic term appear now and then. An often reasonable functional form for a total cost function is
cost 5 b0 1 b1quantity 1 b2quantity2 1 b3quantity3 1 u.
Estimating such a model causes no complications. Interpreting the parameters is more involved
(though straightforward using calculus); we do not study these models further.
6-2c Models with Interaction Terms
Sometimes, it is natural for the partial effect, elasticity, or semi-elasticity of the dependent variable
with respect to an explanatory variable to depend on the magnitude of yet another explanatory variable. For example, in the model
price 5 b0 1 b1sqrft 1 b2bdrms 1 b3sqrft # bdrms 1 b4bthrms 1 u,
the partial effect of bdrms on price (holding all other variables fixed) is
Dprice
5 b2 1 b3sqrft.
Dbdrms
[6.17]
If b3 . 0, then (6.17) implies that an additional bedroom yields a higher increase in housing price for
larger houses. In other words, there is an interaction effect between square footage and number of
bedrooms. In summarizing the effect of bdrms on price, we must evaluate (6.17) at interesting values
58860_ch06_hr_181-219.indd 192
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
193
of sqrft, such as the mean value, or the lower and upper quartiles in the sample. Whether or not b3 is
zero is something we can easily test.
The parameters on the original variables can be tricky to interpret when we include an interaction term. For example, in the previous housing price equation, equation (6.17) shows that b2 is the
effect of bdrms on price for a home with zero square feet! This effect is clearly not of much interest.
Instead, we must be careful to put interesting values of sqrft, such as the mean or median values in the
sample, into the estimated version of equation (6.17).
Often, it is useful to reparameterize a model so that the coefficients on the original variables have
an interesting meaning. Consider a model with two explanatory variables and an interaction:
y 5 b0 1 b1x1 1 b2x2 1 b3x1x2 1 u.
As just mentioned, b2 is the partial effect of x2 on y when x1 5 0. Often, this is not of interest.
Instead, we can reparameterize the model as
y 5 a0 1 d1x1 1 d2x2 1 b3 1 x1 2 m1 2 1 x2 2 m2 2 1 u,
where m1 is the population mean of x1 and m2 is the population mean of x2. We can easily see that
now the coefficient on x2, d2, is the partial effect of x2 on y at the mean value of x1. (By multiplying
out the interaction in the second equation and comparing the coefficients, we can easily show that
d2 5 b2 1 b3m1. The parameter d1 has a similar interpretation.) Therefore, if we subtract the means
of the variables—in practice, these would typically be the sample means—before creating the interaction term, the coefficients on the original variables have a useful interpretation. Plus, we immediately
obtain standard errors for the partial effects at the mean values. Nothing prevents us from replacing
m1 or m2 with other values of the explanatory variables that may be of interest. The following example
illustrates how we can use interaction terms.
Example 6.3
Effects of Attendance on Final Exam Performance
A model to explain the standardized outcome on a final exam (stndfnl) in terms of percentage of
classes attended, prior college grade point average, and ACT score is
stndfnl 5 b0 1 b1atndrte 1 b2priGPA 1 b3ACT 1 b4priGPA2
1 b5ACT2 1 b6priGPA#atndrte 1 u.
[6.18]
(We use the standardized exam score for the reasons discussed in Section 6-1: it is easier to interpret a student’s performance relative to the rest of the class.) In addition to quadratics in priGPA
and ACT, this model includes an interaction between priGPA and the attendance rate. The idea is
that class attendance might have a different effect for students who have performed differently in
the past, as measured by priGPA. We are interested in the effects of attendance on final exam score:
Dstndfnl/Datndrte 5 b1 1 b6priGPA.
Using the 680 observations in ATTEND, for students in a course on microeconomic principles,
the estimated equation is
stndfnl 5 2.05 2 .0067 atndrte 2 1.63 priGPA 2 .128 ACT
1 1.36 2 1 .0102 2
1 .48 2
1 .098 2
1 .296 priGPA2 1 .0045 ACT2 1 .0056 priGPA # atndrte
1 .101 2
1 .0022 2
1 .0043 2
2
2
n 5 680, R 5 .229, R 5 .222.
58860_ch06_hr_181-219.indd 193
[6.19]
10/18/18 4:26 PM
194
PART 1 Regression Analysis with Cross-Sectional Data
We must interpret this equation with extreme care. If we simply look at the coefficient on atndrte, we
will incorrectly conclude that attendance has a negative effect on final exam score. But this coefficient
supposedly measures the effect when priGPA 5 0, which is not interesting (in this sample, the smallest prior GPA is about .86). We must also take care not to look separately at the estimates of b1 and
b6 and conclude that, because each t statistic is insignificant, we cannot reject H0: b1 5 0, b6 5 0.
In fact, the p-value for the F test of this joint hypothesis is .014, so we certainly reject H0 at the 5%
level. This is a good example of where looking at separate t statistics when testing a joint hypothesis
can lead one far astray.
How should we estimate the partial effect of atndrte on stndfnl? We must plug in interesting
values of priGPA to obtain the partial effect. The mean value of priGPA in the sample is 2.59, so at
the mean priGPA, the effect of atndrte on stndfnl is 2.0067 1 .0056 1 2.59 2 < .0078. What does this
mean? Because atndrte is measured as a percentage, it means that a 10 percentage point increase in
atndrte increases stndfnl by .078 standard deviations from the mean final exam score.
How can we tell whether the estimate .0078 is
statistically different from zero? We need to rerun
G o i n g F u rt h e r 6 . 3
the regression, where we replace priGPA·atndrte
If we add the term b7 ACT·atndrte to
with (priGPA − 2.59)·atndrte. This gives, as the
­equation (6.18), what is the partial effect of
new coefficient on atndrte, the estimated effect at
­atndrte on stndfnl?
priGPA 5 2.59, along with its standard error; nothing else in the regression changes. (We described
this device in Section 4.4.) Running this new regression gives the standard error of b^ 1 1 b^ 6 1 2.59 2 5 .0078 as .0026, which yields t 5 .0078/.0026 5 3.
Therefore, at the average priGPA, we conclude that attendance has a statistically significant positive
effect on final exam score.
Things are even more complicated for finding the effect of priGPA on stndfnl because of
the quadratic term priGPA2. To find the effect at the mean value of priGPA and the mean attendance rate, 82, we would replace priGPA2 with 1 priGPA 2 2.59 2 2 and priGPA . atndrte with
priGPA.(atndrte − 82). The coefficient on priGPA becomes the partial effect at the mean values, and
we would have its standard error. (See Computer Exercise C7.)
6-2d Computing Average Partial Effects
The hallmark of models with quadratics, interactions, and other nonlinear functional forms is that the
partial effects depend on the values of one or more explanatory variables. For example, we just saw in
Example 6.3 that the effect of atndrte depends on the value of priGPA. It is easy to see that the partial
effect of priGPA in equation (6.18) is
b2 1 2b4priGPA 1 b6atndrte
(something that can be verified with simple calculus or just by combining the quadratic and interaction formulas). The embellishments in equation (6.18) can be useful for seeing how the strength of
associations between stndfnl and each explanatory variable changes with the values of all explanatory
variables. The flexibility afforded by a model such as (6.18) does have a cost: it is tricky to describe
the partial effects of the explanatory variables on stndfnl with a single number.
Often, one wants a single value to describe the relationship between the dependent variable y and
each explanatory variable. One popular summary measure is the average partial effect (APE), also
called the average marginal effect. The idea behind the APE is simple for models such as (6.18). After
computing the partial effect and plugging in the estimated parameters, we average the partial effects
for each unit across the sample. So, the estimated partial effect of atndrte on stndfnl is
b^ 1 1 b^ 6 priGPAi.
58860_ch06_hr_181-219.indd 194
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
195
We do not want to report this partial effect for each of the 680 students in our sample. Instead, we
average these partial effects to obtain
APEstndfnl 5 b^ 1 1 b^ 6 priGPA,
where priGPA is the sample average of priGPA. The single number APEstndfnl is the (estimated) APE.
The APE of priGPA is only a little more complicated:
APEpriGPA 5 b^ 2 1 2b^ 4priGPA 1 b6atndrte.
Both APEstndfnl and APEpriGPA tell us the size of the partial effects on average.
The centering of explanatory variables about their sample averages before creating quadratics or
interactions forces the coefficient on the levels to be the APEs. This can be cumbersome in complicated models. Fortunately, some commonly used regression packages compute APEs with a simple
command after OLS estimation. Just as importantly, proper standard errors are computed using the
fact that an APE is a linear combination of the OLS coefficients. For example, the APEs and their
standard errors for models with both quadratics and interactions, as in Example 6.3, are easy to obtain.
APEs are also useful in models that are inherently nonlinear in parameters, which we treat in
Chapter 17. At that point we will revisit the definition and calculation of APEs.
6-3 More on Goodness-of-Fit and Selection of Regressors
Until now, we have not focused much on the size of R2 in evaluating our regression models, primarily
because beginning students tend to put too much weight on R-squared. As we will see shortly, choosing a set of explanatory variables based on the size of the R-squared can lead to nonsensical models.
In Chapter 10, we will discover that R-squareds obtained from time series regressions can be artificially high and can result in misleading conclusions.
Nothing about the classical linear model assumptions requires that R2 be above any particular
value; R2 is simply an estimate of how much variation in y is explained by x1, x2, . . . , xk in the population. We have seen several regressions that have had pretty small R-squareds. Although this means
that we have not accounted for several factors that affect y, this does not mean that the factors in u
are correlated with the independent variables. The zero conditional mean assumption MLR.4 is what
determines whether we get unbiased estimators of the ceteris paribus effects of the independent variables, and the size of the R-squared has no direct bearing on this.
A small R-squared does imply that the error variance is large relative to the variance of y, which
means we may have a hard time precisely estimating the bj. But remember, we saw in Section 3.4 that
a large error variance can be offset by a large sample size: if we have enough data, we may be able
to precisely estimate the partial effects even though we have not controlled for many unobserved factors. Whether or not we can get precise enough estimates depends on the application. For example,
suppose that some incoming students at a large university are randomly given grants to buy computer
equipment. If the amount of the grant is truly randomly determined, we can estimate the ceteris paribus effect of the grant amount on subsequent college grade point average by using simple regression
analysis. (Because of random assignment, all of the other factors that affect GPA would be uncorrelated with the amount of the grant.) It seems likely that the grant amount would explain little of the
variation in GPA, so the R-squared from such a regression would probably be very small. But, if we
have a large sample size, we still might get a reasonably precise estimate of the effect of the grant.
Another good illustration of where poor explanatory power has nothing to do with unbiased estimation of the bj is given by analyzing the data set APPLE. Unlike the other data sets we have used,
the key explanatory variables in APPLE were set experimentally—that is, without regard to other
factors that might affect the dependent variable. The variable we would like to explain, ecolbs, is the
(hypothetical) pounds of “ecologically friendly” (“ecolabeled”) apples a family would demand. Each
58860_ch06_hr_181-219.indd 195
10/18/18 4:26 PM
196
PART 1 Regression Analysis with Cross-Sectional Data
family (actually, family head) was presented with a description of ecolabeled apples, along with prices
of regular apples (regprc) and prices of the hypothetical ecolabeled apples (ecoprc). Because the
price pairs were randomly assigned to each family, they are unrelated to other observed factors (such
as family income) and unobserved factors (such as desire for a clean environment). Therefore, the
regression of ecolbs on ecoprc, regprc (across all samples generated in this way) produces u­ nbiased
estimators of the price effects. Nevertheless, the R-squared from the regression is only .0364: the
price variables explain only about 3.6% of the total variation in ecolbs. So, here is a case where
we explain very little of the variation in y, yet we are in the rare situation of knowing that the data
have been generated so that unbiased estimation of the bj is possible. (Incidentally, adding observed
family characteristics has a very small effect on explanatory power. See Computer Exercise C11.)
Remember, though, that the relative change in the R-squared when variables are added to an
equation is very useful: the F statistic in (4.41) for testing the joint significance crucially depends on
the difference in R-squareds between the unrestricted and restricted models.
As we will see in Section 6.4, an important consequence of a low R-squared is that prediction is
difficult. Because most of the variation in y is explained by unobserved factors (or at least factors we
do not include in our model), we will generally have a hard time using the OLS equation to predict
individual future outcomes on y given a set of values for the explanatory variables. In fact, the low
R-squared means that we would have a hard time predicting y even if we knew the bj, the population
coefficients. Fundamentally, most of the factors that explain y are unaccounted for in the explanatory
variables, making prediction difficult.
6-3a Adjusted R-Squared
Most regression packages will report, along with the R-squared, a statistic called the adjusted
R-squared. Because the adjusted R-squared is reported in much applied work, and because it has
some useful features, we cover it in this subsection.
To see how the usual R-squared might be adjusted, it is usefully written as
R2 5 1 2 1 SSR/n 2 / 1 SST/n 2 ,
[6.20]
where SSR is the sum of squared residuals and SST is the total sum of squares; compared with equation (3.28), all we have done is divide both SSR and SST by n. This expression reveals what R2 is actually estimating. Define s2y as the population variance of y and let s2u denote the population variance of
the error term, u. (Until now, we have used s2 to denote s2u, but it is helpful to be more specific here.)
The population R-squared is defined as r2 5 1 2 s2u /s2y ; this is the proportion of the variation in y
in the population explained by the independent variables. This is what R2 is supposed to be estimating.
R2 estimates s2u by SSR/n, which we know to be biased. So why not replace SSR/n with SSR/
(n − k − 1)? Also, we can use SST/(n − 1) in place of SST/n, as the former is the unbiased estimator of
s2y . Using these estimators, we arrive at the adjusted R-squared:
R2 5 1 2 3 SSR/ 1 n 2 k 2 1 2 4 / 3 SST/ 1 n 2 1 2 4
5 1 2 s^ 2 / 3 SST/ 1 n 2 1 2 4 ,
[6.21]
because s^ 2 5 SSR/ 1 n 2 k 2 1 2 . Because of the notation used to denote the adjusted R-squared, it is
sometimes called R-bar squared.
The adjusted R-squared is sometimes called the corrected R-squared, but this is not a good name
because it implies that R2 is somehow better than R2 as an estimator of the population R-squared.
Unfortunately, R2 is not generally known to be a better estimator. It is tempting to think that R2 corrects the bias in R2 for estimating the population R-squared, r2, but it does not: the ratio of two unbiased estimators is not an unbiased estimator.
58860_ch06_hr_181-219.indd 196
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
197
The primary attractiveness of R2 is that it imposes a penalty for adding additional independent
variables to a model. We know that R2 can never fall when a new independent variable is added to a
regression equation: this is because SSR never goes up (and usually falls) as more independent variables are added (assuming we use the same set of observations). But the formula for R2 shows that it
depends explicitly on k, the number of independent variables. If an independent variable is added to
a regression, SSR falls, but so does the df in the regression, n − k − 1. SSR/(n − k − 1) can go up or
down when a new independent variable is added to a regression.
An interesting algebraic fact is the following: if we add a new independent variable to a regression equation, R2 increases if, and only if, the t statistic on the new variable is greater than one in
absolute value. (An extension of this is that R2 increases when a group of variables is added to a
regression if, and only if, the F statistic for joint significance of the new variables is greater than
unity.) Thus, we see immediately that using R2 to decide whether a certain independent variable (or
set of variables) belongs in a model gives us a different answer than standard t or F testing (because a
t or F statistic of unity is not statistically significant at traditional significance levels).
It is sometimes useful to have a formula for R2 in terms of R2. Simple algebra gives
R2 5 1 2 1 1 2 R2 2 1 n 2 1 2 / 1 n 2 k 2 1 2 .
[6.22]
For example, if R2 5 .30, n 5 51, and k 5 10, then R2 5 1 2 .70 1 50 2 /40 5 .125. Thus, for small
n and large k, R2 can be substantially below R2. In fact, if the usual R-squared is small, and n − k − 1
is small, R2 can actually be negative! For example, you can plug in R2 5 .10, n 5 51, and k 5 10
to verify that R2 5 2.125. A negative R2 indicates a very poor model fit relative to the number of
degrees of freedom.
The adjusted R-squared is sometimes reported along with the usual R-squared in regressions, and
sometimes R2 is reported in place of R2. It is important to remember that it is R2, not R2, that appears
in the F statistic in (4.41). The same formula with R2r and R2ur is not valid.
6-3b Using Adjusted R-Squared to Choose between
Nonnested Models
In Section 4-5, we learned how to compute an F statistic for testing the joint significance of a group
of variables; this allows us to decide, at a particular significance level, whether at least one variable in
the group affects the dependent variable. This test does not allow us to decide which of the variables
has an effect. In some cases, we want to choose a model without redundant independent variables,
and the adjusted R-squared can help with this.
In the major league baseball salary example in Section 4-5, we saw that neither hrunsyr nor
rbisyr was individually significant. These two variables are highly correlated, so we might want to
choose between the models
and
log 1 salary 2 5 b0 1 b1years 1 b2gamesyr 1 b3bavg 1 b4hrunsyr 1 u
log 1 salary 2 5 b0 1 b1years 1 b2gamesyr 1 b3bavg 1 b4rbisyr 1 u.
These two equations are nonnested models because neither equation is a special case of the other.
The F statistics we studied in Chapter 4 only allow us to test nested models: one model (the restricted
model) is a special case of the other model (the unrestricted model). See equations (4.32) and (4.28)
for examples of restricted and unrestricted models. One possibility is to create a composite model
that contains all explanatory variables from the original models and then to test each model against
the general model using the F test. The problem with this process is that either both models might
58860_ch06_hr_181-219.indd 197
10/18/18 4:26 PM
198
PART 1 Regression Analysis with Cross-Sectional Data
be rejected or neither model might be rejected (as happens with the major league baseball salary
example in Section 4-5). Thus, it does not always provide a way to distinguish between models with
nonnested regressors.
In the baseball player salary regression using the data in MLB1, R2 for the regression containing
hrunsyr is .6211, and R2 for the regression containing rbisyr is .6226. Thus, based on the adjusted
R-squared, there is a very slight preference for the model with rbisyr. But the difference is practically very small, and we might obtain a different answer by controlling for some of the variables in
Computer Exercise C5 in Chapter 4. (Because both nonnested models contain five parameters, the
usual R-squared can be used to draw the same conclusion.)
Comparing R2 to choose among different nonnested sets of independent variables can be valuable when these variables represent different functional forms. Consider two models relating R&D
intensity to firm sales:
rdintens 5 b0 1 b1log 1 sales 2 1 u.
rdintens 5 b0 1 b1sales 1 b2sales2 1 u.
[6.23]
[6.24]
The first model captures a diminishing return by including sales in logarithmic form; the second
model does this by using a quadratic. Thus, the second model contains one more parameter than
the first.
When equation (6.23) is estimated using the 32 observations on chemical firms in RDCHEM, R2
is .061, and R2 for equation (6.24) is .148. Therefore, it appears that the quadratic fits much better. But
a comparison of the usual R-squareds is unfair to the first model because it contains one fewer parameter than (6.24). That is, (6.23) is a more parsimonious model than (6.24).
Everything else being equal, simpler models are better. Because the usual R-squared does not
­penalize more complicated models, it is better to use R2. The R2 for (6.23) is .030, while R2 for (6.24)
is .090. Thus, even after adjusting for the difference in degrees of freedom, the quadratic model wins
out. The quadratic model is also preferred when profit margin is added to each regression.
There is an important limitation in using R2 to choose between nonnested models: we cannot
use it to choose between different functional forms for the dependent variable. This is unfortunate,
because we often want to decide on whether y or log(y) (or maybe some other transformation) should
be used as the dependent variable based on goodness-of-fit. But neither R2 nor R2 can be used for
this purpose. The reason is simple: these R-squareds measure the explained proportion of the total
variation in whatever dependent variable we are using in the regression, and different nonlinear functions of the dependent variable will have different amounts of variation to explain. For example, the
total variations in y and log(y) are not the same and
G o i n g F u rt h e r 6 . 4
are often very different. Comparing the adjusted
R-squareds from regressions with these different
Explain why choosing a model by maximiz2
forms of the dependent variables does not tell us
ing R or minimizing s^ (the standard error of
anything about which model fits better; they are fitthe regression) is the same thing.
ting two separate dependent variables.
Example 6.4
CEO Compensation and Firm Performance
Consider two estimated models relating CEO compensation to firm performance:
salary 5 830.63 1 .0163 sales 1 19.63 roe
1 223.90 2 1 .0089 2
1 11.08 2
n 5 209, R2 5 .029, R2 5 .020
58860_ch06_hr_181-219.indd 198
[6.25]
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
199
and
lsalary 5 4 .36 1 .275 lsales 1 .0179 roe
1 0.29 2 1 .033 2
1 .0040 2[6.26]
n 5 209, R2 5 .282, R2 5 .275,
where roe is the return on equity discussed in Chapter 2. For simplicity, lsalary and lsales denote the
natural logs of salary and sales. We already know how to interpret these different estimated equations. But can we say that one model fits better than the other?
The R-squared for equation (6.25) shows that sales and roe explain only about 2.9% of the variation in CEO salary in the sample. Both sales and roe have marginal statistical significance.
Equation (6.26) shows that log(sales) and roe explain about 28.2% of the variation in log(salary).
In terms of goodness-of-fit, this much higher R-squared would seem to imply that model (6.26) is
much better, but this is not necessarily the case. The total sum of squares for salary in the sample
is 391,732,982, while the total sum of squares for log(salary) is only 66.72. Thus, there is much less
variation in log(salary) that needs to be explained.
At this point, we can use features other than R2 or R2 to decide between these models. For example, log(sales) and roe are much more statistically significant in (6.26) than are sales and roe in (6.25),
and the coefficients in (6.26) are probably of more interest. To be sure, however, we will need to make
a valid goodness-of-fit comparison.
In Section 6-4, we will offer a goodness-of-fit measure that does allow us to compare models
where y appears in both level and log form.
6-3c Controlling for Too Many Factors in Regression Analysis
In many of the examples we have covered, and certainly in our discussion of omitted variables bias in
Chapter 3, we have worried about omitting important factors from a model that might be correlated with
the independent variables. It is also possible to control for too many variables in a regression analysis.
If we overemphasize goodness-of-fit, we open ourselves to controlling for factors in a regression
model that should not be controlled for. To avoid this mistake, we need to remember the ceteris paribus interpretation of multiple regression models.
To illustrate this issue, suppose we are doing a study to assess the impact of state beer taxes on
traffic fatalities. The idea is that a higher tax on beer will reduce alcohol consumption, and likewise
drunk driving, resulting in fewer traffic fatalities. To measure the ceteris paribus effect of taxes on
fatalities, we can model fatalities as a function of several factors, including the beer tax:
fatalities 5 b0 1 b1tax 1 b2miles 1 b3percmale 1 b4perc16_21 1 p ,
where
miles 5 total miles driven.
percmale 5 percentage of the state population that is male.
perc16_21 5 percentage of the population between ages 16 and 21, and so on.
Notice how we have not included a variable measuring per capita beer consumption. Are we
committing an omitted variables error? The answer is no. If we control for beer consumption in this
equation, then how would beer taxes affect traffic fatalities? In the equation
fatalities 5 b0 1 b1tax 1 b2beercons 1 p ,
58860_ch06_hr_181-219.indd 199
10/18/18 4:26 PM
200
PART 1 Regression Analysis with Cross-Sectional Data
b1 measures the difference in fatalities due to a one percentage point increase in tax, holding beercons
fixed. It is difficult to understand why this would be interesting. We should not be controlling for differences in beercons across states, unless we want to test for some sort of indirect effect of beer taxes.
Other factors, such as gender and age distribution, should be controlled for.
As a second example, suppose that, for a developing country, we want to estimate the effects of
pesticide usage among farmers on family health expenditures. In addition to pesticide usage amounts,
should we include the number of doctor visits as an explanatory variable? No. Health expenditures
include doctor visits, and we would like to pick up all effects of pesticide use on health expenditures.
If we include the number of doctor visits as an explanatory variable, then we are only measuring the
effects of pesticide use on health expenditures other than doctor visits. It makes more sense to use
number of doctor visits as a dependent variable in a separate regression on pesticide amounts.
The previous examples are what can be called over controlling for factors in multiple regression.
Often this results from nervousness about potential biases that might arise by leaving out an important
explanatory variable. But it is important to remember the ceteris paribus nature of multiple regression.
In some cases, it makes no sense to hold some factors fixed precisely because they should be allowed
to change when a policy variable changes.
Unfortunately, the issue of whether or not to control for certain factors is not always clear-cut.
For example, Betts (1995) studies the effect of high school quality on subsequent earnings. He points
out that, if better school quality results in more education, then controlling for education in the regression along with measures of quality will underestimate the return to quality. Betts does the analysis
with and without years of education in the equation to get a range of estimated effects for quality of
schooling.
To see explicitly how pursuing high R-squareds can lead to trouble, consider the housing price
example from Section 4-5 that illustrates the testing of multiple hypotheses. In that case, we wanted to
test the rationality of housing price assessments. We regressed log(price) on log(assess), log(lotsize),
log(sqrft), and bdrms and tested whether the latter three variables had zero population coefficients
while log(assess) had a coefficient of unity. But what if we change the purpose of the analysis and
estimate a hedonic price model, which allows us to obtain the marginal values of various housing
attributes? Should we include log(assess) in the equation? The adjusted R-squared from the regression with log(assess) is .762, while the adjusted R-squared without it is .630. Based on goodnessof-fit only, we should include log(assess). But this is incorrect if our goal is to determine the effects
of lot size, square footage, and number of bedrooms on housing values. Including log(assess) in the
equation amounts to holding one measure of value fixed and then asking how much an additional
bedroom would change another measure of value. This makes no sense for valuing housing attributes.
If we remember that different models serve different purposes, and we focus on the ceteris paribus interpretation of regression, then we will not include the wrong factors in a regression model.
6-3d Adding Regressors to Reduce the Error Variance
We have just seen some examples of where certain independent variables should not be included in
a regression model, even though they are correlated with the dependent variable. From Chapter 3,
we know that adding a new independent variable to a regression can exacerbate the multicollinearity
problem. On the other hand, because we are taking something out of the error term, adding a variable
generally reduces the error variance. Generally, we cannot know which effect will dominate.
However, there is one case that is clear: we should always include independent variables that
affect y and are uncorrelated with all of the independent variables of interest. Why? Because adding
such a variable does not induce multicollinearity in the population (and therefore multicollinearity in
the sample should be negligible), but it will reduce the error variance. In large sample sizes, the standard errors of all OLS estimators will be reduced.
As an example, consider estimating the individual demand for beer as a function of the average
county beer price. It may be reasonable to assume that individual characteristics are uncorrelated with
58860_ch06_hr_181-219.indd 200
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
201
county-level prices, and so a simple regression of beer consumption on county price would suffice for
estimating the effect of price on individual demand. But it is possible to get a more precise estimate
of the price elasticity of beer demand by including individual characteristics, such as age and amount
of education. If these factors affect demand and are uncorrelated with price, then the standard error of
the price coefficient will be smaller, at least in large samples.
As a second example, consider the grants for computer equipment given at the beginning of
Section 6-3. If, in addition to the grant variable, we control for other factors that can explain college
GPA, we can probably get a more precise estimate of the effect of the grant. Measures of high school
grade point average and rank, SAT and ACT scores, and family background variables are good candidates. Because the grant amounts are randomly assigned, all additional control variables are uncorrelated with the grant amount; in the sample, multicollinearity between the grant amount and other
independent variables should be minimal. But adding the extra controls might significantly reduce
the error variance, leading to a more precise estimate of the grant effect. Remember, the issue is not
unbiasedness here: we obtain an unbiased and consistent estimator whether or not we add the high
school performance and family background variables. The issue is getting an estimator with a smaller
sampling variance.
A related point is that when we have random assignment of a policy, we need not worry about
whether some of our explanatory variables are “endogenous”—provided these variables themselves
are not affected by the policy. For example, in studying the effect of hours in a job training program
on labor earnings, we can include the amount of education reported prior to the job training program.
We need not worry that schooling might be correlated with omitted factors, such as “ability,” because
we are not trying to estimate the return to schooling. We are trying to estimate the effect of the job
training program, and we can include any controls that are not themselves affected by job training
without biasing the job training effect. What we must avoid is including a variable such as the amount
of education after the job training program, as some people may decide to get more education because
of how many hours they were assigned to the job training program.
Unfortunately, cases where we have information on additional explanatory variables that are
uncorrelated with the explanatory variables of interest are somewhat rare in the social sciences. But
it is worth remembering that when these variables are available, they can be included in a model to
reduce the error variance without inducing multicollinearity.
6-4 Prediction and Residual Analysis
In Chapter 3, we defined the OLS predicted or fitted values and the OLS residuals. Predictions are
certainly useful, but they are subject to sampling variation, because they are obtained using the OLS
estimators. Thus, in this section, we show how to obtain confidence intervals for a prediction from the
OLS regression line.
From Chapters 3 and 4, we know that the residuals are used to obtain the sum of squared residuals and the R-squared, so they are important for goodness-of-fit and testing. Sometimes, economists
study the residuals for particular observations to learn about individuals (or firms, houses, etc.) in the
sample.
6.4a Confidence Intervals for Predictions
Suppose we have estimated the equation
y^ 5 b^ 0 1 b^ 1x1 1 b^ 2x2 1 p 1 b^ kxk.
[6.27]
When we plug in particular values of the independent variables, we obtain a prediction for y, which
is an estimate of the expected value of y given the particular values for the explanatory variables.
For emphasis, let c1, c2, . . . , ck denote particular values for each of the k independent variables; these
58860_ch06_hr_181-219.indd 201
10/18/18 4:26 PM
202
PART 1 Regression Analysis with Cross-Sectional Data
may or may not correspond to an actual data point in our sample. The parameter we would like to
estimate is
The estimator of u 0 is
u 0 5 b0 1 b1c1 1 b2c2 1 p 1 bkck
5 E 1 y 0 x1 5 c1, x2 5 c2, . . . , xk 5 ck 2 .
u^ 0 5 b^ 0 1 b^ 1c1 1 b^ 2c2 1 p 1 bkck.
[6.28]
[6.29]
In practice, this is easy to compute. But what if we want some measure of the uncertainty in this predicted value? It is natural to construct a confidence interval for u 0, which is centered at u^ 0.
To obtain a confidence interval for u 0, we need a standard error for u^ 0. Then, with a large df, we
can construct a 95% confidence interval using the rule of thumb u^ 0 6 2 # se 1 u^ 0 2 . (As always, we can
use the exact percentiles in a t distribution.)
How do we obtain the standard error of u^ 0 ? This is the same problem we encountered in
Section 4-4: we need to obtain a standard error for a linear combination of the OLS estimators. Here,
the problem is even more complicated, because all of the OLS estimators generally appear in u^ 0
(unless some cj are zero). Nevertheless, the same trick that we used in Section 4-4 will work here.
Write b0 5 u 0 2 b1c1 2 p 2 bkck and plug this into the equation
y 5 b0 1 b1x1 1 p 1 bkxk 1 u
to obtain
y 5 u 0 1 b1 1 x1 2 c1 2 1 b2 1 x2 2 c2 2 1 p 1 bk 1 xk 2 ck 2 1 u.
[6.30]
yi on 1 xi1 2 c1 2 , . . . , 1 xik 2 ck 2 , i 5 1, 2, . . . , n.
[6.31]
In other words, we subtract the value cj from each observation on xj, and then we run the regression of
The predicted value in (6.29) and, more importantly, its standard error, are obtained from the intercept
(or constant) in regression (6.31).
As an example, we obtain a confidence interval for a prediction from a college GPA regression,
where we use high school information.
Example 6.5
Confidence Interval for Predicted College GPA
Using the data in GPA2, we obtain the following equation for predicting college GPA:
colgpa 5 1.493 1 .00149 sat 2 .01386 hsperc
1 0.075 2 1 .00007 2
1 .00056 2
2 .06088 hsize 1 .00546 hsize2
1 .01650 2
1 .00227 2 [6.32]
n 5 4,137, R2 5 .278, R2 5 .277, s^ 5 .560,
where we have reported estimates to several digits to reduce round-off error. What is predicted college GPA, when sat 5 1,200, hsperc 5 30, and hsize 5 5 (which means 500)? This is easy to get by
plugging these values into equation (6.32): colgpa 5 2.70 (rounded to two digits). Unfortunately, we
cannot use equation (6.32) directly to get a confidence interval for the expected colgpa at the given
58860_ch06_hr_181-219.indd 202
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
203
values of the independent variables. One simple way to obtain a confidence interval is to define a new
set of independent variables: sat0 5 sat 2 1,200, hsperc0 5 hsperc 2 30, hsize0 5 hsize 2 5, and
hsizesq0 5 hsize2 2 25. When we regress colgpa on these new independent variables, we get
colgpa 5 2.700 1 .00149 sat0 2 .01386 hsperc0
1 0.020 2 1 .00007 2
1 .00056 2
2 .06088 hsize0 1 .00546 hsizesq0
1 .01650 2
1 .00227 2
n 5 4,137, R2 5 .278, R2 5 .277, s^ 5 .560.
The only difference between this regression and that in (6.32) is the intercept, which is the prediction we want, along with its standard error, .020. It is not an accident that the slope coefficients, their
standard errors, R-squared, and so on are the same as before; this provides a way to check that the
proper transformations were done. We can easily construct a 95% confidence interval for the expected
college GPA: 2.70 6 1.96(.020) or about 2.66 to 2.74. This confidence interval is rather narrow due to
the very large sample size.
Because the variance of the intercept estimator is smallest when each explanatory variable has
zero sample mean (see Problem 10, part (iv) in Chapter 2 for the simple regression case), it follows
from the regression in (6.31) that the variance of the prediction is smallest at the mean values of the xj.
(That is, cj 5 xj for all j.) This result is not too surprising, as we have the most faith in our regression
line near the middle of the data. As the values of the cj get farther away from the xj, Var 1 y^ 2 gets larger
and larger.
The previous method allows us to put a confidence interval around the OLS estimate of
E 1 y 0 x1, p , xk 2 for any values of the explanatory variables. In other words, we obtain a confidence
interval for the average value of y for the subpopulation with a given set of covariates. But a confidence interval for the average person in the subpopulation is not the same as a confidence interval for
a particular unit (individual, family, firm, and so on) from the population. In forming a confidence
interval for an unknown outcome on y, we must account for another very important source of variation: the variance in the unobserved error, which measures our ignorance of the unobserved factors
that affect y.
Let y0 denote the value for which we would like to construct a confidence interval, which we
sometimes call a prediction interval. For example, y0 could represent a person or firm not in our
original sample. Let x01, p, x0k be the new values of the independent variables, which we assume we
observe, and let u0 be the unobserved error. Therefore, we have
y0 5 b0 1 b1x01 1 b2x02 1 p 1 bkx0k 1 u0.
[6.33]
As before, our best prediction of y0 is the expected value of y0 given the explanatory variables, which
we estimate from the OLS regression line: y^ 0 5 b^ 0 1 b^ 1x01 1 b^ 2x02 1 p 1 b^ kx0k . The prediction ­error
in using y^ 0 to predict y0 is
e^ 0 5 y0 2 y^ 0 5 1 b0 1 b1x01 1 p 1 bkx0k 2 1 u0 2 y^ 0.
[6.34]
Now, E 1 y^ 0 2 5 E 1 b^ 0 2 1 E 1 b^ 1 2 x01 1 E 1 b^ 2 2 x02 1 p 1 E 1 b^ k 2 x0k 5 b0 1 b1x01 1 p 1 bkx0k , because
the b^ j are unbiased. (As before, these expectations are all conditional on the sample values of the
­independent variables.) Because u0 has zero mean, E 1 e^ 0 2 5 0. We have shown that the expected
­prediction error is zero.
In finding the variance of e^ 0, note that u0 is uncorrelated with each b^ j, because u0 is uncorrelated with
the errors in the sample used to obtain the b^ j. By basic properties of covariance (see Math Refresher B),
58860_ch06_hr_181-219.indd 203
10/18/18 4:26 PM
204
PART 1 Regression Analysis with Cross-Sectional Data
u0 and y^ 0 are uncorrelated. Therefore, the variance of the prediction error (conditional on all
in-sample values of the independent variables) is the sum of the variances:
Var 1 e^ 0 2 5 Var 1 y^ 0 2 1 Var 1 u0 2 5 Var 1 y^ 0 2 1 s2,
[6.35]
se 1 e^ 0 2 5 5 3 se 1 y^ 0 2 4 2 1 s^ 2 6 1/ 2.
[6.36]
where s2 5 Var 1 u0 2 is the error variance. There are two sources of variation in e^ 0. The first is the
sampling error in y^ 0, which arises because we have estimated the bj. Because each b^ j has a variance proportional to 1/n, where n is the sample size, Var 1 y^ 0 2 is proportional to 1/n. This means
that, for large samples, Var 1 y^ 0 2 can be very small. By contrast, s2 is the variance of the error in
the population; it does not change with the sample size. In many examples, s2 will be the dominant term in (6.35).
Under the classical linear model assumptions, the b^ j and u0 are normally distributed, and so e^ 0 is
also normally distributed (conditional on all sample values of the explanatory variables). Earlier, we
described how to obtain an unbiased estimator of Var 1 y^ 0 2 , and we obtained our unbiased estimator of
s2 in Chapter 3. By using these estimators, we can define the standard error of e^ 0 as
Using the same reasoning for the t statistics of the b^ j, e^ 0/se 1 e^ 0 2 has a t distribution with n − (k 1 1)
degrees of freedom. Therefore,
P 3 2t.025 # e^ 0/se 1 e^ 0 2 # t.025 4 5 .95,
where t.025 is the 97.5th percentile in the tn2k21 distribution. For large n − k − 1, remember that
t.025 < 1.96. Plugging in e^ 0 5 y0 2 y^ 0 and rearranging gives a 95% prediction interval for y0:
y^ 0 6 t.025 # se 1 e^ 0 2 ;
[6.37]
as usual, except for small df, a good rule of thumb is y^ 0 6 2se 1 e^ 0 2 . This is wider than the confidence
interval for y^ 0 itself because of s^ 2 in (6.36); it often is much wider to reflect the factors in u0 that we
have not accounted for.
Example 6.6
Confidence Interval for Future College GPA
Suppose we want a 95% confidence interval for the future college GPA of a high school student with
sat 5 1,200, hsperc 5 30, and hsize 5 5. In Example 6.5, we obtained a 95% CI for the average
college GPA among all students with the particular characteristics sat 5 1,200, hsperc 5 30, and
hsize 5 5. Now, we want a 95% CI for any particular student with these characteristics. The 95%
prediction interval must account for the variation in the individual, unobserved characteristics that
affect college performance. We have everything we need to obtain a CI for colgpa. se 1 y^ 0 2 5 .020 and
s^ 5 .560 and so, from (6.36), se 1 e^ 0 2 5 3 1 .020 2 2 1 1 .560 2 2 4 1/ 2 < .560. Notice how small se 1 y^ 0 2 is
relative to s^ : virtually all of the variation in e^ 0 comes from the variation in u0. The 95% CI is 2.70 6
1.96(.560) or about 1.60 to 3.80. This is a wide confidence interval and shows that, based on the factors we included in the regression, we cannot accurately pin down an individual’s future college grade
point average. (In one sense, this is good news, as it means that high school rank and performance
on the SAT do not preordain one’s performance in college.) Evidently, the unobserved characteristics
that affect college GPA vary widely among individuals with the same observed SAT score and high
school rank.
58860_ch06_hr_181-219.indd 204
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
205
6-4b Residual Analysis
Sometimes, it is useful to examine individual observations to see whether the actual value of the
dependent variable is above or below the predicted value; that is, to examine the residuals for the
individual observations. This process is called residual analysis. Economists have been known to
examine the residuals from a regression in order to aid in the purchase of a home. The following
housing price example illustrates residual analysis. Housing price is related to various observable
characteristics of the house. We can list all of the characteristics that we find important, such as size,
number of bedrooms, number of bathrooms, and so on. We can use a sample of houses to estimate a
relationship between price and attributes, where we end up with a predicted value and an actual value
for each house. Then, we can construct the residuals, u^ i 5 yi 2 y^ i. The house with the most negative
residual is, at least based on the factors we have controlled for, the most underpriced one relative to its
observed characteristics. Of course, a selling price substantially below its predicted price could indicate some undesirable feature of the house that we have failed to account for, and which is therefore
contained in the unobserved error. In addition to obtaining the prediction and residual, it also makes
sense to compute a confidence interval for what the future selling price of the home could be, using
the method described in equation (6.37).
Using the data in HPRICE1, we run a regression of price on lotsize, sqrft, and bdrms. In the
sample of 88 homes, the most negative residual is 2120.206, for the 81st house. Therefore, the asking
price for this house is $120,206 below its predicted price.
There are many other uses of residual analysis. One way to rank law schools is to regress median
starting salary on a variety of student characteristics (such as median LSAT scores of entering class,
median college GPA of entering class, and so on) and to obtain a predicted value and residual for
each law school. The law school with the largest residual has the highest predicted value added. (Of
course, there is still much uncertainty about how an individual’s starting salary would compare with
the median for a law school overall.) These residuals can be used along with the costs of attending
each law school to determine the best value; this would require an appropriate discounting of future
earnings.
Residual analysis also plays a role in legal decisions. A New York Times article entitled “Judge
Says Pupil’s Poverty, Not Segregation, Hurts Scores” (6/28/95) describes an important legal case.
The issue was whether the poor performance on standardized tests in the Hartford School District,
relative to performance in surrounding suburbs, was due to poor school quality at the highly segregated schools. The judge concluded that “the disparity in test scores does not indicate that Hartford
is doing an inadequate or poor job in educating its students or that its schools are failing, because the
predicted scores based upon the relevant socioeconomic factors are about at the levels that one would
expect.” This conclusion is based on a regression analysis of average or median scores on socioeconomic characteristics of various school districts in
Connecticut. The judge’s conclusion suggests that,
G o i n g F u rt h e r 6 . 5
given the poverty levels of students at Hartford
schools, the actual test scores were similar to those
How would you use residual analysis to
predicted from a regression analysis: the residual
determine which professional athletes
for Hartford was not sufficiently negative to conare overpaid or underpaid relative to their
clude that the schools themselves were the cause of
performance?
low test scores.
6-4c Predicting y When log(y) Is the Dependent Variable
Because the natural log transformation is used so often for the dependent variable in empirical economics, we devote this subsection to the issue of predicting y when log(y) is the dependent variable.
As a byproduct, we will obtain a goodness-of-fit measure for the log model that can be compared with
the R-squared from the level model.
58860_ch06_hr_181-219.indd 205
10/18/18 4:26 PM
206
PART 1 Regression Analysis with Cross-Sectional Data
To obtain a prediction, it is useful to define logy 5 log 1 y 2 ; this emphasizes that it is the log of y
that is predicted in the model
logy 5 b0 1 b1x1 1 b2x2 1 p 1 bkxk 1 u.
[6.38]
In this equation, the xj might be transformations of other variables; for example, we could have
x1 5 log 1 sales 2 , x2 5 log 1 mktval 2 , x3 5 ceoten in the CEO salary example.
Given the OLS estimators, we know how to predict logy for any value of the independent variables:
logy 5 b^ 0 1 b^ 1x1 1 b^ 2x2 1 p 1 b^ kxk.
[6.39]
Now, because the exponential undoes the log, our first guess for predicting y is to simply exponentiate the predicted value for log 1 y 2 : y^ 5 exp 1 logy 2 . This does not work; in fact, it will systematically
underestimate the expected value of y. In fact, if model (6.38) follows the CLM assumptions MLR.1
through MLR.6, it can be shown that
E 1 y 0 x 2 5 exp 1 s2/2 2 # exp 1 b0 1 b1x1 1 b2x2 1 p 1 bkxk 2 ,
where x denotes the independent variables and s2 is the variance of u. 3 If u , Normal 1 0,s2 2 , then the
expected value of exp(u) is exp 1 s2 /2 2 . 4 This equation shows that a simple adjustment is needed to
predict y:
y^ 5 exp 1 s^ 2/2 2 exp 1 logy 2 ,
[6.40]
E 1 y 0 x 2 5 a0exp 1 b0 1 b1x1 1 b2x2 1 p 1 bkxk 2 ,
[6.41]
y^ 5 a^ 0exp 1 logy 2 ,
[6.42]
where s^ 2 is simply the unbiased estimator of s2. Because s^ , the standard error of the regression, is
always reported, obtaining predicted values for y is easy. Because s^ 2 . 0, exp 1 s^ 2/2 2 . 1. For large
s^ 2, this adjustment factor can be substantially larger than unity.
The prediction in (6.40) is not unbiased, but it is consistent. There are no unbiased predictions of
y, and in many cases, (6.40) works well. However, it does rely on the normality of the error term, u.
In Chapter 5, we showed that OLS has desirable properties, even when u is not normally distributed.
Therefore, it is useful to have a prediction that does not rely on normality. If we just assume that u is
independent of the explanatory variables, then we have
where a0 is the expected value of exp(u), which must be greater than unity.
Given an estimate a^ 0, we can predict y as
which again simply requires exponentiating the predicted value from the log model and multiplying
the result by a^ 0.
Two approaches suggest themselves for estimating a0 without the normality assumption. The first is
based on a0 5 E 3 exp 1 u 2 4. To estimate a0 we replace the population expectation with a sample average and
then we replace the unobserved errors, ui, with the OLS residuals, u^ i 5 log 1 yi 2 2 b^ 0 2 b^ 1xi1 2 p 2 b^ kxik.
This leads to the method of moments estimator (see Math Refresher C)
a^ 0 5 n21 a exp 1 u^ i 2 .
n
[6.43]
i51
Not surprisingly, a^ 0 is a consistent estimator of a0, but it is not unbiased because we have replaced
ui with u^ i inside a nonlinear function. This version of a^ 0 is a special case of what Duan (1983) called
a smearing estimate. Because the OLS residuals have a zero sample average, it can be shown that,
for any data set, a^ 0 . 1. (Technically, a^ 0 would equal one if all the OLS residuals were zero, but this
58860_ch06_hr_181-219.indd 206
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
207
never happens in any interesting application.) That a^ 0 is necessarily greater than one is convenient
because it must be that a0 . 1.
A different estimate of a0 is based on a simple regression through the origin. To see how it works,
define mi 5 exp 1 b0 1 b1xi1 1 p 1 bkxik 2 , so that, from equation (6.41), E 1 yi 0 mi 2 5 a0mi. If we
could observe the mi, we could obtain an unbiased estimator of a0 from the regression yi on mi without
an intercept. Instead, we replace the bj with their OLS estimates and obtain m^ i 5 exp 1 logyi 2 , where,
of course, the logyi are the fitted values from the regression logyi on xi1, . . . , xik (with an intercept).
Then ǎ0 [to distinguish it from a^ 0 in equation (6.43)] is the OLS slope estimate from the simple
regression yi on m^ i (no intercept):
ǎ0 5 a a m^ 2i b
n
i51
21
a a m^ iyi b.
n
[6.44]
i51
We will call ǎ0 the regression estimate of a0. Like a^ 0, ǎ0 is consistent but not unbiased. Interestingly,
ǎ0 is not guaranteed to be greater than one, although it will be in most applications. If ǎ0 is less than
one, and especially if it is much less than one, it is likely that the assumption of independence between
u and the xj is violated. If ǎ0 , 1, one possibility is to just use the estimate in (6.43), although this
may simply be masking a problem with the linear model for log(y).
We summarize the steps:
6-4d Predicting y When the Dependent Variable Is log(y)
1.
2.
3.
4.
Obtain the fitted values, logyi, and residuals, u^ i, from the regression logy on x1, . . . , xk.
Obtain a^ 0 as in equation (6.43) or ǎ0 in equation (6.44).
For given values of x1, . . . , xk, obtain logy from (6.42).
Obtain the prediction y^ from (6.42) (with a^ 0 or ǎ0).
We now show how to predict CEO salaries using this procedure.
Example 6.7
Predicting CEO Salaries
The model of interest is
log 1 salary 2 5 b0 1 b1log 1 sales 2 1 b2log 1 mktval 2 1 b3ceoten 1 u,
so that b1 and b2 are elasticities and 100 # b3 is a semi-elasticity. The estimated equation using
CEOSAL2 is
lsalary 5 4.504 1 .163 lsales 1 .109 lmktval 1 .0117 ceoten
1 .257 2 1 .039 2
1 .050 2
1 .0053 2
[6.45]
2
n 5 177, R 5 .318,
where, for clarity, we let lsalary denote the log of salary, and similarly for lsales and lmktval. Next,
we obtain m^ i 5 exp 1 lsalaryi 2 for each observation in the sample.
The Duan smearing estimate from (6.43) is about a^ 0 5 1.136, and the regression estimate from
(6.44) is ǎ0 5 1.117. We can use either estimate to predict salary for any values of sales, mktval, and
ceoten. Let us find the prediction for sales 5 5,000 (which means $5 billion because sales is in millions), mktval 5 10,000 (or $10 billion), and ceoten 5 10. From (6.45), the prediction for lsalary is
4.504 1 .163 # log 1 5,000 2 1 .109 # log 1 10.000 2 1 .0117 1 10 2 < 7.013, and exp 1 7.013 2 < 1,110.983.
Using the estimate of a0 from (6.43), the predicted salary is about 1,262.077, or $1,262,077. Using
the estimate from (6.44) gives an estimated salary of about $1,240,968. These differ from each other
by much less than each differs from the naive prediction of $1,110,983.
58860_ch06_hr_181-219.indd 207
10/18/18 4:26 PM
208
PART 1 Regression Analysis with Cross-Sectional Data
We can use the previous method of obtaining predictions to determine how well the model with
log(y) as the dependent variable explains y. We already have measures for models when y is the
dependent variable: the R-squared and the adjusted R-squared. The goal is to find a goodness-of-fit
measure in the log(y) model that can be compared with an R-squared from a model where y is the
dependent variable.
There are different ways to define a goodness-of-fit measure after retransforming a model for
log(y) to predict y. Here we present two approaches that are easy to implement. The first gives the
same goodness-of-fit measures whether we estimate a0 as in (6.40), (6.43), or (6.44). To motivate the
measure, recall that in the linear regression equation estimated by OLS,
y^ 5 b^ 0 1 b^ 1x1 1 p 1 b^ kxk,
[6.46]
the usual R-squared is simply the square of the correlation between yi and y^ i (see Section 3-2). Now, if
instead we compute fitted values from (6.42)—that is, y^ i 5 a^ 0mi for all observations i—then it makes
sense to use the square of the correlation between yi and these fitted values as an R-squared. Because
correlation is unaffected if we multiply by a constant, it does not matter which estimate of a0 we use.
In fact, this R-squared measure for y [not log(y)] is just the squared correlation between yi and m^ i. We
can compare this directly with the R-squared from equation (6.46).
The squared correlation measure does not depend on how we estimate a0. A second approach is
to compute an R-squared for y based on a sum of squared residuals. For concreteness, suppose we use
equation (6.43) to estimate a0. Then the residual for predicting yi is
r^ i 5 yi 2 a^ 0 exp 1 logyi 2 ,
[6.47]
and we can use these residuals to compute a sum of squared residuals. Using the formula for R-squared
from linear regression, we are led to
g i51r^ 2i
n
g i51 1 yi 2 y 2
n
12
[6.48]
as an alternative goodness-of-fit measure that can be compared with the R-squared from the linear
model for y. Notice that we can compute such a measure for the alternative estimates of a0 in equation
(6.40) and (6.44) by inserting those estimates in place of a^ 0 in (6.47). Unlike the squared correlation
between yi and m^ i, the R-squared in (6.48) will depend on how we estimate a0. The estimate that minin
mizes g i51r^ 2i is that in equation (6.44), but that does not mean we should prefer it (and certainly not if
ǎ0 , 1). We are not really trying to choose among the different estimates of a0; rather, we are finding
goodness-of-fit measures that can be compared with the linear model for y.
Example 6.8
Predicting CEO Salaries
After we obtain the m^ i, we just obtain the correlation between salaryi and m^ i; it is .493. The square of
it is about .243, and this is a measure of how well the log model explains the variation in salary, not
log(salary). [The R2 from (6.45), .318, tells us that the log model explains about 31.8% of the variation in log(salary).]
As a competing linear model, suppose we estimate a model with all variables in levels:
salary 5 b0 1 b1sales 1 b2mktval 1 b3ceoten 1 u.
[6.49]
The key is that the dependent variable is salary. We could use logs of sales or mktval on the
right-hand side, but it makes more sense to have all dollar values in levels if one (salary) appears as
a level. The R-squared from estimating this equation using the same 177 observations is .201. Thus,
the log model explains more of the variation in salary, and so we prefer it to (6.49) on goodness-of-fit
58860_ch06_hr_181-219.indd 208
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
209
grounds. The log model is also preferred because it seems more realistic and its parameters are easier
to interpret.
If we maintain the full set of classical linear model assumptions in the model (6.38), we can easily obtain prediction intervals for y0 5 exp 1 b0 1 b1x01 1 p 1 bkx0k 1 u0 2 when we have estimated a
linear model for log(y). Recall that x01, x02, . . . , x0k are known values and u0 is the unobserved error that
partly determines y0. From equation (6.37), a 95% prediction interval for logy0 5 log 1 y0 2 is simply
logy0 6 t.025 # se 1 e^ 0 2 , where se 1 e^ 0 2 is obtained from the regression of log(y) on x1, . . . , xk using the
original n observations. Let cl 5 logy0 2t.025 # se 1 e^ 0 2 and cu 5 logy0 1 t.025 # se 1 e^ 0 2 be the lower and
upper bounds of the prediction interval for logy0. That is, P 1 cl # logy0 # cu 2 5 .95. Because the exponential function is strictly increasing, it is also true that P 3 exp 1 cl 2 # exp 1 logy0 2 # exp 1 cu 2 4 5 .95,
that is, P 3 exp 1 cl 2 # y0 # exp 1 cu 2 4 5 .95. Therefore, we can take exp 1 cl 2 and exp 1 cu 2 as the lower and
upper bounds, respectively, for a 95% prediction interval for y0. For large n, t.025 5 1.96, and so a 95%
prediction interval for y0 is exp 3 21.96 # se 1 e^ 0 2 4 exp 1 b^ 0 1 x0b^ 2 to exp 3 21.96 # se 1 e^ 0 2 4 exp 1 b^ 0 1 x0b^ 2 ,
where x0b^ is shorthand for b^ 1x01 1 p 1 b^ kx0k . Remember, the b^ j and se 1 e^ 0 2 are obtained from the
regression with log(y) as the dependent variable. Because we assume normality of u in (6.38), we
probably would use (6.40) to obtain a point prediction for y0. Unlike in equation (6.37), this point prediction will not lie halfway between the lower and upper bounds exp 1 cl 2 and exp 1 cu 2 . One can obtain
different 95% prediction intervalues by choosing different quantiles in the tn2k21 distribution. If qa1
and qa2 are quantiles with a2 2 a1 5 .95, then we can choose cl 5 qa1se 1 e^ 0 2 and cu 5 qa2se 1 e^ 0 2 .
As an example, consider the CEO salary regression, where we make the prediction at the
same values of sales, mktval, and ceoten as in Example 6.7. The standard error of the regression
for (6.43) is about .505, and the standard error of logy0 is about .075. Therefore, using equation
(6.36), se 1 e^ 0 2 < .511; as in the GPA example, the error variance swamps the estimation error in the
parameters, even though here the sample size is only 177. A 95% prediction interval for salary0 is
exp 3 21.96 # 1 .511 2 4 exp 1 7.013 2 to exp 3 1.96 # 1 .511 2 4 exp 1 7.013 2 , or about 408.071 to 3,024.678,
that is, $408,071 to $3,024,678. This very wide 95% prediction interval for CEO salary at the given
sales, market value, and tenure values shows that there is much else that we have not included in the
regression that determines salary. Incidentally, the point prediction for salary, using (6.40), is about
$1,262,075—higher than the predictions using the other estimates of a0 and closer to the lower bound
than the upper bound of the 95% prediction interval.
Summary
In this chapter, we have covered some important multiple regression analysis topics.
Section 6-1 showed that a change in the units of measurement of an independent variable changes
the OLS coefficient in the expected manner: if xj is multiplied by c, its coefficient is divided by c. If the
dependent variable is multiplied by c, all OLS coefficients are multiplied by c. Neither t nor F statistics are
affected by changing the units of measurement of any variables.
We discussed beta coefficients, which measure the effects of the independent variables on the dependent variable in standard deviation units. The beta coefficients are obtained from a standard OLS regression
after the dependent and independent variables have been transformed into z-scores.
We provided a detailed discussion of functional form, including the logarithmic transformation, quadratics, and interaction terms. It is helpful to summarize some of our conclusions.
Considerations When Using Logarithms
1. The coefficients have percentage change interpretations. We can be ignorant of the units of measurement of any variable that appears in logarithmic form, and changing units from, say, dollars to thousands of dollars has no effect on a variable’s coefficient when that variable appears in logarithmic
form.
58860_ch06_hr_181-219.indd 209
10/18/18 4:26 PM
210
PART 1 Regression Analysis with Cross-Sectional Data
2. Logs are often used for dollar amounts that are always positive, as well as for variables such as population, especially when there is a lot of variation. They are used less often for variables measured in
years, such as schooling, age, and experience. Logs are used infrequently for variables that are already
percents or proportions, such as an unemployment rate or a pass rate on a test.
3. Models with log(y) as the dependent variable often more closely satisfy the classical linear model
assumptions. For example, the model has a better chance of being linear, homoskedasticity is more
likely to hold, and normality is often more plausible.
4. In many cases, taking the log greatly reduces the variation of a variable, making OLS estimates less
prone to outlier influence. However, in cases where y is a fraction and close to zero for many observations, log 1 yi 2 can have much more variability than yi. For values yi very close to zero, log 1 yi 2 is a negative number very large in magnitude.
5. If y $ 0 but y 5 0 is possible, we cannot use log(y). Sometimes log(1 1 y) is used, but interpretation
of the coefficients is difficult.
6. For large changes in an explanatory variable, we can compute a more accurate estimate of the percentage change effect.
7. It is harder (but possible) to predict y when we have estimated a model for log(y).
Considerations When Using Quadratics
1. A quadratic function in an explanatory variable allows for an increasing or decreasing effect.
2. The turning point of a quadratic is easily calculated, and it should be calculated to see if it makes sense.
3. Quadratic functions where the coefficients have the opposite sign have a strictly positive turning point;
if the signs of the coefficients are the same, the turning point is at a negative value of x.
4. A seemingly small coefficient on the square of a variable can be practically important in what it
implies about a changing slope. One can use a t test to see if the quadratic is statistically significant,
and compute the slope at various values of x to see if it is practically important.
5. For a model quadratic in a variable x, the coefficient on x measures the partial effect starting from x 5 0,
as can be seen in equation (6.11). If zero is not a possible or interesting value of x, one can center x
about a more interesting value, such as the average in the sample, before computing the square. This is
the same as computing the average partial effect. Computing Exercise C12 provides an example.
Considerations When Using Interactions
1. Interaction terms allow the partial effect of an explanatory variable, say x1, to depend on the level of
another variable, say x2—and vice versa.
2. Interpreting models with interactions can be tricky. The coefficient on x1, say b1, measures the partial
effect of x1 on y when x2 5 0, which may be impossible or uninteresting. Centering x1 and x2 around
interesting values before constructing the interaction term typically leads to an equation that is visually
more appealing. When the variables are centered about their sample averages before multiplying them
together to create the interaction, the coefficients on the levels become estimated average partial effects.
3. A standard t test can be used to determine if an interaction term is statistically significant. Computing
the partial effects at different values of the explanatory variables can be used to determine the practical
importance of interactions.
We introduced the adjusted R-squared, R2, as an alternative to the usual R-squared for measuring goodness-of-fit. Whereas R2 can never fall when another variable is added to a regression, R2 penalizes the number of regressors and can drop when an independent variable is added. This makes R2 preferable for choosing
between nonnested models with different numbers of explanatory variables. Neither R2 nor R2 can be used to
compare models with different dependent variables. Nevertheless, it is fairly easy to obtain g­ oodness-of-fit
measures for choosing between y and log(y) as the dependent variable, as shown in Section 6-4.
58860_ch06_hr_181-219.indd 210
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
211
In Section 6-3, we discussed the somewhat subtle problem of relying too much on R2 or R2 in arriving
at a final model: it is possible to control for too many factors in a regression model. For this reason, it is
important to think ahead about model specification, particularly the ceteris paribus nature of the multiple
regression equation. Explanatory variables that affect y and are uncorrelated with all the other explanatory
variables can be used to reduce the error variance without inducing multicollinearity.
In Section 6-4, we demonstrated how to obtain a confidence interval for a prediction made from an
OLS regression line. We also showed how a confidence interval can be constructed for a future, unknown
value of y.
Occasionally, we want to predict y when log(y) is used as the dependent variable in a regression
model. Section 6-4 explains this simple method. Finally, we are sometimes interested in knowing about the
sign and magnitude of the residuals for particular observations. Residual analysis can be used to determine
whether particular members of the sample have predicted values that are well above or well below the
actual outcomes.
Key Terms
Adjusted R-Squared
Average Partial Effect (APE)
Beta Coefficients
Bootstrap
Bootstrap Standard Error
Interaction Effect
Quadratic Functions
Resampling Method
Residual Analysis
Smearing Estimate
Standardized Coefficients
Variance of the Prediction Error
Nonnested Models
Over Controlling
Population R-Squared
Prediction Error
Prediction Interval
Predictions
Problems
1 The following equation was estimated using the data in CEOSAL1:
log 1 salary 2 5 4.322 1 .276 log 1 sales 2 1 .0215 roe 2 .00008 roe2
1 .324 2 1 .033 2
1 .0129 2
1 .00026 2
2
n 5 209, R 5 .282.
This equation allows roe to have a diminishing effect on log(salary). Is this generality necessary?
Explain why or why not.
^ 1, . . . , b^ k be the OLS estimates from the regression of yi on xi1, . . . , xik, i 5 1, 2, . . . , n. For
2 Let b^ 0, b
nonzero constants c1, . . . , ck, argue that the OLS intercept and slopes from the regression of c0yi on
| 5 c b^ , |b 5 1 c /c 2 b^ , . . . , b^ 5 1 c /c 2 b^ . [Hint:
c1xi1, . . . , ckxik, i 5 1, 2, . . . , n, are given by b
0
0 0
1
0 1
1
k
0 k
k
| must solve the first order
Use the fact that the b^ j solve the first order conditions in (3.13), and the b
j
conditions involving the rescaled dependent and independent variables.]
3 Using the data in RDCHEM, the following equation was obtained by OLS:
rdintens 5 2.613 1 .00030 sales 2 .0000000070 sales2
1 .429 2 1 .00014 2
2
n 5 32, R 5 .1484.
(i)
(ii)
58860_ch06_hr_181-219.indd 211
1 .0000000037 2
At what point does the marginal effect of sales on rdintens become negative?
Would you keep the quadratic term in the model? Explain.
10/18/18 4:26 PM
212
PART 1 Regression Analysis with Cross-Sectional Data
(iii) Define salesbil as sales measured in billions of dollars: salesbil 5 sales/1,000. Rewrite the
estimated equation with salesbil and salesbil2 as the independent variables. Be sure to report
standard errors and the R-squared. [Hint: Note that salesbil2 5 sales2/ 1 1,000 2 2.]
(iv) For the purpose of reporting the results, which equation do you prefer?
4 The following model allows the return to education to depend upon the total amount of both parents’
education, called pareduc:
(i)
log 1 wage 2 5 b0 1 b1educ 1 b2educ # pareduc 1 b3exper 1 b4tenure 1 u.
Show that, in decimal form, the return to another year of education in this model is
Dlog 1 wage 2 /Deduc 5 b1 1 b2 pareduc.
(ii)
What sign do you expect for b2? Why?
Using the data in WAGE2, the estimated equation is
log 1 wage 2 5 5.65 1 .047 educ 1 .00078 educ # pareduc 1
1 .13 2 1 .010 2
1 .00021 2
.019 exper 1 .010 tenure
1 .004 2
1 .003 2
n 5 722, R2 5 .169.
(Only 722 observations contain full information on parents’ education.) Interpret the coefficient
on the interaction term. It might help to choose two specific values for pareduc—for example,
pareduc 5 32 if both parents have a college education, or pareduc 5 24 if both parents have a
high school education—and to compare the estimated return to educ.
(iii) When pareduc is added as a separate variable to the equation, we get:
log 1 wage 2 5 4.94 1 .097 educ 1 .033 pareduc 2 .0016 educ # pareduc
1 .38 2 1 .027 2
1 .017 2
1 .0012 2
1 .020 exper 1 .010 tenure
1 .004 2
1 .003 2
2
n 5 722, R 5 .174.
Does the estimated return to education now depend positively on parent education? Test the null
hypothesis that the return to education does not depend on parent education.
5 In Example 4.2, where the percentage of students receiving a passing score on a tenth-grade math
exam (math10) is the dependent variable, does it make sense to include sci11—the percentage of eleventh graders passing a science exam—as an additional explanatory variable?
6 When atndrte2 and ACT # atndrte are added to the equation estimated in (6.19), the R-squared becomes
.232. Are these additional terms jointly significant at the 10% level? Would you include them in the
model?
7 The following three equations were estimated using the 1,534 observations in 401K:
prate 5 80.29 1 5.44 mrate 1 .269 age 2 .00013 totemp
1 .78 2 1 .52 2
1 .045 2
1 .00004 2
2
2
R 5 .100, R 5 .098.
58860_ch06_hr_181-219.indd 212
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
213
prate 5 97.32 1 5.02 mrate 1 .314 age 2 2.66 log 1 totemp 2
2
1 1.95 2 1 0.51 2
2
R 5 .144, R 5 .142.
1 .044 2
1 .28 2
prate 5 80.62 1 5.34 mrate 1 .290 age 2 .00043 totemp
1 .78 2 1 .52 2
1 .045 2
1 .00009 2
2
1 .0000000039 totemp
1 .00000000010 2
R2 5 .108, R2 5 .106.
Which of these three models do you prefer? Why?
8 Suppose we want to estimate the effects of alcohol consumption (alcohol) on college grade point average (colGPA). In addition to collecting information on grade point averages and alcohol usage, we also
obtain attendance information (say, percentage of lectures attended, called attend). A standardized test
score (say, SAT) and high school GPA (hsGPA) are also available.
(i) Should we include attend along with alcohol as explanatory variables in a multiple regression
model? (Think about how you would interpret balcohol.)
(ii) Should SAT and hsGPA be included as explanatory variables? Explain.
9 If we start with (6.38) under the CLM assumptions, assume large n, and ignore the estimation e­ rror
in the b^ j , a 95% prediction interval for y0 is 3 exp 1 21.96s^ 2 exp 1 logy0 2 , exp 1 1.96s^ 2 exp 1 logy0 2 4 .
The point prediction for y0 is y^ 0 5 exp 1 s^ 2/2 2 exp 1 logy0 2 .
(i) For what values of s^ will the point prediction be in the 95% prediction interval? Does this condition seem likely to hold in most applications?
(ii) Verify that the condition from part (i) is satisfied in the CEO salary example.
10 The following two equations were estimated using the data in MEAPSINGLE. The key explanatory
variable is lexppp, the log of expenditures per student at the school level.
math4 5 24.49 1 9.01 lexppp 2 .422 free 2 .752 lmedinc 2 .274 pctsgle
1 59.24 2 1 4.04 2
1 .071 2
1 5.358 2
1 .161 2
n 5 229, R2 5 .472, R2 5 .462.
math4 5 149.38 1 1.93 lexppp 2 .060 free 2 10.78 lmedinc 2 .397 pctsgle 1 .667 read4
1 41.70 2 1 2.82 2
1 .054 2
1 3.76 2
1 .111 2
1 .042 2
2
2
n 5 229, R 5 .749, R 5 .743.
(i)
If you are a policy maker trying to estimate the causal effect of per-student spending on math
test performance, explain why the first equation is more relevant than the second. What is the
estimated effect of a 10% increase in expenditures per student?
(ii) Does adding read4 to the regression have strange effects on coefficients and statistical significance other than blexppp?
(iii) How would you explain to someone with only basic knowledge of regression why, in this case,
you prefer the equation with the smaller adjusted R-squared?
11 Consider the equation
y 5 b0 1 b1x 1 b2x2 1 u
E 1 u 0 x 2 5 0,
where the explanatory variable x has a standard normal distribution in the population. In particular,
E(x) 5 0, E(x2) 5 Var(x) 5 1, and E(x3) 5 0. This last condition holds because the standard normal
58860_ch06_hr_181-219.indd 213
10/18/18 4:26 PM
214
PART 1 Regression Analysis with Cross-Sectional Data
distribution is symmetric about zero. We want to study what we can say about the OLS estimator if we
omit x2 and compute the simple regression estimator of the intercept and slope.
(i) Show that we can write
y 5 a0 1 b1x 1 v.
where E 1 v 2 5 0. In particular, find v and the new intercept, a0.
(ii) Show that E 1 v 0 x 2 depends on x unless b2 5 0.
(iii) Show that Cov(x, v) = 0.
(iv) If b^ 1 is the slope coefficient from regression yi on xi, is b^ 1 consistent for b1? Is it unbiased?
Explain.
(v) Argue that being able to estimate b1 has some value in the following sense: b1 is the partial
­effect of x on E 1 y 0 x 2 evaluated at x = 0, the average value of x.
(vi) Explain why being able to consistently estimate b1 and b2 is more valuable than just estimating b1.
Computer Exercises
C1 Use the data in KIELMC, only for the year 1981, to answer the following questions. The data are for
houses that sold during 1981 in North Andover, Massachusetts; 1981 was the year construction began
on a local garbage incinerator.
(i) To study the effects of the incinerator location on housing price, consider the simple regression
model
log 1 price 2 5 b0 1 b1log 1 dist 2 1 u,
where price is housing price in dollars and dist is distance from the house to the ­incinerator
measured in feet. Interpreting this equation causally, what sign do you expect for b1 if the
presence of the incinerator depresses housing prices? Estimate this equation and interpret the
results.
(ii) To the simple regression model in part (i), add the variables log(intst), log(area), log(land),
rooms, baths, and age, where intst is distance from the home to the interstate, area is square
footage of the house, land is the lot size in square feet, rooms is total number of rooms, baths is
number of bathrooms, and age is age of the house in years. Now, what do you conclude about
the effects of the incinerator? Explain why (i) and (ii) give conflicting results.
(iii) Add 3 log 1 intst 2 4 2 to the model from part (ii). Now what happens? What do you conclude about
the importance of functional form?
(iv) Is the square of log(dist) significant when you add it to the model from part (iii)?
C2 Use the data in WAGE1 for this exercise.
(i) Use OLS to estimate the equation
log 1 wage 2 5 b0 1 b1educ 1 b2exper 1 b3exper2 1 u
and report the results using the usual format.
(ii) Is exper2 statistically significant at the 1% level?
(iii) Using the approximation
%Dwage < 100 1 b^ 2 1 2b^ 3exper 2 Dexper,
(iv)
58860_ch06_hr_181-219.indd 214
find the approximate return to the fifth year of experience. What is the approximate return to the
twentieth year of experience?
At what value of exper does additional experience actually lower predicted log(wage)? How
many people have more experience in this sample?
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
215
C3 Consider a model in which the return to education depends upon the amount of work experience (and
vice versa):
(i)
log 1 wage 2 5 b0 1 b1educ 1 b2exper 1 b3educ # exper 1 u.
Show that the return to another year of education (in decimal form), holding exper fixed, is
b1 1 b3exper.
(ii) State the null hypothesis that the return to education does not depend on the level of exper.
What do you think is the appropriate alternative?
(iii) Use the data in WAGE2 to test the null hypothesis in (ii) against your stated alternative.
(iv) Let u 1 denote the return to education (in decimal form), when exper 5 10: u 1 5 b1 1 10b3.
Obtain u^ 1 and a 95% confidence interval for u 1. (Hint: Write b1 5 u 1 2 10b3 and plug this
into the equation; then rearrange. This gives the regression for obtaining the confidence interval
for u 1.)
C4 Use the data in GPA2 for this exercise.
(i) Estimate the model
sat 5 b0 1 b1hsize 1 b2hsize2 1 u,
where hsize is the size of the graduating class (in hundreds), and write the results in the usual
form. Is the quadratic term statistically significant?
(ii) Using the estimated equation from part (i), what is the “optimal” high school size? Justify your
answer.
(iii) Is this analysis representative of the academic performance of all high school seniors?
Explain.
(iv) Find the estimated optimal high school size, using log(sat) as the dependent variable. Is it much
different from what you obtained in part (ii)?
C5 Use the housing price data in HPRICE1 for this exercise.
(i) Estimate the model
log 1 price 2 5 b0 1 b1log 1 lotsize 2 1 b2log 1 sqrft 2 1 b3bdrms 1 u
and report the results in the usual OLS format.
Find the predicted value of log(price), when lotsize 5 20,000, sqrft 5 2,500, and bdrms 5 4.
Using the methods in Section 6-4, find the predicted value of price at the same values of the
explanatory variables.
(iii) For explaining variation in price, decide whether you prefer the model from part (i) or the
model
(ii)
price 5 b0 1 b1lotsize 1 b2sqrft 1 b3bdrms 1 u.
C6 Use the data in VOTE1 for this exercise.
(i) Consider a model with an interaction between expenditures:
voteA 5 b0 1 b1prtystrA 1 b2expendA 1 b3expendB 1 b4expendA # expendB 1 u.
What is the partial effect of expendB on voteA, holding prtystrA and expendA fixed? What is the
partial effect of expendA on voteA? Is the expected sign for b4 obvious?
(ii) Estimate the equation in part (i) and report the results in the usual form. Is the interaction term
statistically significant?
(iii) Find the average of expendA in the sample. Fix expendA at 300 (for $300,000). What is the estimated effect of another $100,000 spent by Candidate B on voteA? Is this a large effect?
58860_ch06_hr_181-219.indd 215
10/18/18 4:26 PM
216
PART 1 Regression Analysis with Cross-Sectional Data
(iv)
Now fix expendB at 100. What is the estimated effect of DexpendA 5 100 on voteA? Does this
make sense?
(v) Now, estimate a model that replaces the interaction with shareA, Candidate A’s percentage share
of total campaign expenditures. Does it make sense to hold both expendA and expendB fixed,
while changing shareA?
(vi) (Requires calculus) In the model from part (v), find the partial effect of expendB on voteA, holding prtystrA and expendA fixed. Evaluate this at expendA 5 300 and expendB 5 0 and comment on the results.
C7 Use the data in ATTEND for this exercise.
(i) In the model of Example 6.3, argue that
Dstndfnl/DpriGPA < b2 1 2b4 priGPA 1 b6atndrte.
(ii)
Use equation (6.19) to estimate the partial effect when priGPA 5 2.59 and atndrte 5 82.
Interpret your estimate.
Show that the equation can be written as
stndfnl 5 u 0 1 b1atndrte 1 u 2 priGPA 1 b3 ACT 1 b4 1 priGPA 2 2.59 2 2
1 b5ACT2 1 b6 priGPA 1 atndrte 2 82 2 1 u,
where u 2 5 b2 1 2b4 1 2.59 2 1 b6 1 82 2 . (Note that the intercept has changed, but this is unimportant.) Use this to obtain the standard error of u^ 2 from part (i).
(iii) Suppose that, in place of priGPA(atndrte − 82), you put 1 priGPA 2 2.59 2 ? (atndrte − 82).
Now how do you interpret the coefficients on atndrte and priGPA?
C8 Use the data in HPRICE1 for this exercise.
(i) Estimate the model
price 5 b0 1 b1lotsize 1 b2sqrft 1 b3bdrms 1 u
and report the results in the usual form, including the standard error of the regression. Obtain
predicted price, when we plug in lotsize 5 10,000, sqrft 5 2,300, and bdrms 5 4; round this
price to the nearest dollar.
(ii) Run a regression that allows you to put a 95% confidence interval around the predicted value in
part (i). Note that your prediction will differ somewhat due to rounding error.
(iii) Let price0 be the unknown future selling price of the house with the characteristics used in parts
(i) and (ii). Find a 95% CI for price0 and comment on the width of this confidence interval.
C9 The data set NBASAL contains salary information and career statistics for 269 players in the National
Basketball Association (NBA).
(i) Estimate a model relating points-per-game (points) to years in the league (exper), age, and years
played in college (coll). Include a quadratic in exper; the other variables should appear in level
form. Report the results in the usual way.
(ii) Holding college years and age fixed, at what value of experience does the next year of experience actually reduce points-per-game? Does this make sense?
(iii) Why do you think coll has a negative and statistically significant coefficient? (Hint: NBA p­ layers
can be drafted before finishing their college careers and even directly out of high school.)
(iv) Add a quadratic in age to the equation. Is it needed? What does this appear to imply about the
effects of age, once experience and education are controlled for?
(v) Now regress log(wage) on points, exper, exper2, age, and coll. Report the results in the usual
format.
58860_ch06_hr_181-219.indd 216
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
217
(vi) Test whether age and coll are jointly significant in the regression from part (v). What does this
imply about whether age and education have separate effects on wage, once productivity and
seniority are accounted for?
C10 Use the data in BWGHT2 for this exercise.
(i) Estimate the equation
log 1 bwght 2 5 b0 1 b1npvis 1 b2npvis2 1 u
by OLS, and report the results in the usual way. Is the quadratic term significant?
Show that, based on the equation from part (i), the number of prenatal visits that maximizes
log(bwght) is estimated to be about 22. How many women had at least 22 prenatal visits in the
sample?
(iii) Does it make sense that birth weight is actually predicted to decline after 22 prenatal visits? Explain.
(iv) Add mother’s age to the equation, using a quadratic functional form. Holding npvis fixed, at
what mother’s age is the birth weight of the child maximized? What fraction of women in the
sample are older than the “optimal” age?
(v) Would you say that mother’s age and number of prenatal visits explain a lot of the variation in
log(bwght)?
(vi) Using quadratics for both npvis and age, decide whether using the natural log or the level of
bwght is better for predicting bwght.
(ii)
C11 Use APPLE to verify some of the claims made in Section 6-3.
(i) Run the regression ecolbs on ecoprc, regprc and report the results in the usual form, ­including
the R-squared and adjusted R-squared. Interpret the coefficients on the price variables and
­comment on their signs and magnitudes.
(ii) Are the price variables statistically significant? Report the p-values for the individual t tests.
(iii) What is the range of fitted values for ecolbs? What fraction of the sample reports ecolbs 5 0?
Comment.
(iv) Do you think the price variables together do a good job of explaining variation in ecolbs?
Explain.
(v) Add the variables faminc, hhsize (household size), educ, and age to the regression from part (i).
Find the p-value for their joint significance. What do you conclude?
(vi) Run separate simple regressions of ecolbs on ecoprc and then ecolbs on regprc. How do the
simple regression coefficients compare with the multiple regression from part (i)? Find the
­correlation coefficient between ecoprc and regprc to help explain your findings.
(vii) Consider a model that adds family income and the quantity demanded for regular apples:
ecolbs 5 b0 1 b1ecoprc 1 b2regprc 1 b3 faminc 1 b4reglbs 1 u.
From basic economic theory, which explanatory variable does not belong to the equation? When
you drop the variables one at a time, do the sizes of the adjusted R-squareds affect your answer?
C12 Use the subset of 401KSUBS with fsize 5 1; this restricts the analysis to single-person households;
see also Computer Exercise C8 in Chapter 4.
(i) The youngest age in the sample is 25. How many people are 25 years old?
(ii) In the model
nettfa 5 b0 1 b1inc 1 b2age 1 b3age2 1 u,
what is the literal interpretation of b2? By itself, is it of much interest?
(iii) Estimate the model from part (ii) and report the results in standard form. Are you concerned that
the coefficient on age is negative? Explain.
58860_ch06_hr_181-219.indd 217
10/18/18 4:26 PM
218
PART 1 Regression Analysis with Cross-Sectional Data
(iv)
(v)
Because the youngest people in the sample are 25, it makes sense to think that, for a given
level of income, the lowest average amount of net total financial assets is at age 25. Recall
that the partial effect of age on nettfa is b2 1 2b3age, so the partial effect at age 25 is
b2 1 2b3 1 25 2 5 b2 1 50b3; call this u 2. Find u^ 2 and obtain the two-sided p-value for testing
H0: u 2 5 0. You should conclude that u^ 2 is small and very statistically insignificant. [Hint: One
way to do this is to estimate the model nettfa 5 a0 1 b1inc 1 u 2age 1 b3 1 age 2 25 2 2 1 u,
where the intercept, a0 is different from b0. There are other ways, too.]
Because the evidence against H0: u 2 5 0 is very weak, set it to zero and estimate the model
nettfa 5 a0 1 b1inc 1 b3 1 age 2 25 2 2 1 u.
In terms of goodness-of-fit, does this model fit better than that in part (ii)?
(vi) For the estimated equation in part (v), set inc 5 30 (roughly, the average value) and graph the
relationship between nettfa and age, but only for age $ 25. Describe what you see.
(vii) Check to see whether including a quadratic in inc is necessary.
C13 Use the data in MEAP00 to answer this question.
(i) Estimate the model
math4 5 b0 1 b2lexppp 1 b2lenroll 1 b3lunch 1 u
by OLS, and report the results in the usual form. Is each explanatory variable statistically
­significant at the 5% level?
(ii) Obtain the fitted values from the regression in part (i). What is the range of fitted values? How
does it compare with the range of the actual data on math4?
(iii) Obtain the residuals from the regression in part (i). What is the building code of the school that
has the largest (positive) residual? Provide an interpretation of this residual.
(iv) Add quadratics of all explanatory variables to the equation, and test them for joint significance.
Would you leave them in the model?
(v) Returning to the model in part (i), divide the dependent variable and each explanatory variable
by its sample standard deviation, and rerun the regression. (Include an intercept unless you also
first subtract the mean from each variable.) In terms of standard deviation units, which explanatory variable has the largest effect on the math pass rate?
C14 Use the data in BENEFITS to answer this question. It is a school-level data set at the K–5 level on average teacher salary and benefits. See Example 4.10 for background.
(i) Regress lavgsal on bs and report the results in the usual form. Can you reject
H0: bbs 5 0 against a two-sided alternative? Can you reject H0: bbs 5 21 against
H1: bbs . 21? Report the p-values for both tests.
(ii) Define lbs 5 log 1 bs 2 . Find the range of values for lbs and find its standard deviation.
How do these compare to the range and standard deviation for bs?
(iii) Regress lavgsal on lbs. Does this fit better than the regression from part (i)?
(iv) Estimate the equation
lavgsal 5 b0 1 b1bs 1 b2 lenroll 1 b3lstaff 1 b4lunch 1 u
and report the results in the usual form. What happens to the coefficient on bs? Is it now statistically different from zero?
(v) Interpret the coefficient on lstaff. Why do you think it is negative?
(vi) Add lunch2 to the equation from part (iv). Is it statistically significant? Compute the turning
point (minimum value) in the quadratic, and show that it is within the range of the observed data
on lunch. How many values of lunch are higher than the calculated turning point?
(vii) Based on the findings from part (vi), describe how teacher salaries relate to school poverty rates.
In terms of teacher salary, and holding other factors fixed, is it better to teach at a school with
lunch 5 0 (no poverty), lunch 5 50, or lunch 5 100 (all kids eligible for the free lunch program)?
58860_ch06_hr_181-219.indd 218
10/18/18 4:26 PM
CHAPTER 6 Multiple Regression Analysis: Further Issues
219
Appendix 6A
6A. A Brief Introduction to Bootstrapping
In many cases where formulas for standard errors are hard to obtain mathematically, or where they
are thought not to be very good approximations to the true sampling variation of an estimator, we
can rely on a resampling method. The general idea is to treat the observed data as a population
that we can draw samples from. The most common resampling method is the bootstrap. (There are
actually several versions of the bootstrap, but the most general, and most easily applied, is called the
nonparametric bootstrap, and that is what we describe here.)
Suppose we have an estimate, u^ , of a population parameter, u. We obtained this estimate, which
could be a function of OLS estimates (or estimates that we cover in later chapters), from a random
sample of size n. We would like to obtain a standard error for u^ that can be used for constructing t
statistics or confidence intervals. Remarkably, we can obtain a valid standard error by computing the
estimate from different random samples drawn from the original data.
Implementation is easy. If we list our observations from 1 through n, we draw n numbers randomly, with replacement, from this list. This produces a new data set (of size n) that consists of the
original data, but with many observations appearing multiple times (except in the rather unusual case
that we resample the original data). Each time we randomly sample from the original data, we can
estimate u using the same procedure that we used on the original data. Let u^ 1b2 denote the estimate
from bootstrap sample b. Now, if we repeat the resampling and estimation m times, we have m new
estimates, 5 u^ 1b2: b 5 1, 2, p , m 6 . The bootstrap standard error of u^ is just the sample standard
deviation of the u^ 1b2, namely,
bse 1 u^ 2 5 c 1 m 2 1 2 21 a 1 u^ 1b2 2 u^ 2 2 d
m
1/ 2
,
[6.50]
b51
where u^ is the average of the bootstrap estimates.
If obtaining an estimate of u on a sample of size n requires little computational time, as in the
case of OLS and all the other estimators we encounter in this text, we can afford to choose m—the
number of bootstrap replications—to be large. A typical value is m 5 1,000, but even m 5 500 or
a somewhat smaller value can produce a reliable standard error. Note that the size of m—the number of times we resample the original data—has nothing to do with the sample size, n. (For certain
estimation problems beyond the scope of this text, a large n can force one to do fewer bootstrap
replications.) Many statistics and econometrics packages have built-in bootstrap commands, and this
makes the calculation of bootstrap standard errors simple, especially compared with the work often
required to obtain an analytical formula for an asymptotic standard error.
One can actually do better in most cases by using the bootstrap sample to compute p-values for
t statistics (and F statistics), or for obtaining confidence intervals, rather than obtaining a bootstrap
standard error to be used in the construction of t statistics or confidence intervals. See Horowitz
(2001) for a comprehensive treatment.
58860_ch06_hr_181-219.indd 219
10/18/18 4:26 PM
chapter
7
Multiple Regression
Analysis with Qualitative
Information
A
lmost all of our discussion in the previous chapters has focused on the case where the dependent and independent variables in our multiple regression models have quantitative meaning.
Just a few examples include hourly wage rate, years of education, college grade point average,
amount of air pollution, level of firm sales, and number of arrests. In each case, the magnitude of the
variable conveys useful information. In some cases, we take the natural log and then the coefficients
can be turned into percentage changes.
In Section 2-7 we introduced the notion of a binary (or dummy) explanatory variable, and we discussed how simple regression on a binary variable can be used to evaluate randomized interventions.
We showed how to extend program evaluation to the multiple regression case in Sections 3-7e and 4-7
when it is necessary to account for observed differences between the control and treatment groups.
The purpose of this chapter is to provide a comprehensive analysis of how to include qualitative factors into regression models. In addition to indicators of participating in a program, or being
subjected to a new policy, the race or ethnicity of an individual, marital status, the industry of a firm
(manufacturing, retail, and so on), and the region in the United States where a city is located (South,
North, West, and so on) are common examples of qualitative factors.
After we discuss the appropriate ways to describe qualitative informationinSection7-1, we show
how qualitative explanatory variables can be easily incorporated into multiple regression models
in Sections 7-2, 7-3, and 7-4. These sections cover almost all of the popular ways that qualitative
220
58860_ch07_hr_220-261.indd 220
10/18/18 4:27 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
221
independent variables are used in cross-sectional regression analysis, including creating interactions
among qualitative variables and between qualitative and quantitative variables.
In Section 7-5, we discuss the case where our dependent variable is binary, which is a particular
kind of qualitative dependent variable. The multiple regression model is called the linear probability
model (LPM), and the coefficients can be interpreted as changes in a probability. While much maligned
by some econometricians, the simplicity of the LPM makes it useful in many empirical contexts. We
will describe drawbacks of the LPM in Section 7-5, but they are often secondary in empirical work.
Section 7.6 reconsiders policy analysis, including the potential outcomes perspective, and proposes a flexible regression approach for estimating the effects of interventions. Section 7.7 is a short
section that explains how to interpret multiple regression estimates when y is a discrete variable that
has quantitative meeting.
This chapter does not assume you have read the material on potential outcomes and policy analysis in Chapters 2, 3, and 4, and so it stands alone as a dicussion of how to incorporate qualitative
information into regression.
7-1 Describing Qualitative Information
Qualitative factors often come in the form of binary information: a person is female or male; a person
does or does not own a personal computer; a firm offers a certain kind of employee pension plan or it
does not; a state administers capital punishment or it does not. In all of these examples, the relevant information can be captured by defining a binary variable or a zero-one variable. In econometrics, binary
variables are most commonly called dummy variables, although this name is not especially descriptive.
In defining a dummy variable, we must decide
which event is assigned the value one and which is
G o i n g F u rt h e r 7 . 1
assigned the value zero. For example, in a study of indiSuppose that, in a study comparing elecvidual wage determination, we might define female to
tion outcomes between Democratic and
be a binary variable taking on the value one for females
Republican candidates, you wish to indicate
and the value zero for males. The name in this case
the party of each candidate. Is a name such
indicates the event with the value one. The same inforas party a wise choice for a binary variable
mation is captured by defining male to be one if the
in this case? What would be a better name?
person is male and zero if the person is female. Either
of these is better than using gender because this name
does not make it clear when the dummy variable is one: does gender 5 1 correspond to male or female?
What we call our variables is unimportant for getting regression results, but it always helps to choose
names that clarify equations and expositions.
Suppose in the wage example that we have chosen the name female to indicate gender. Further,
we define a binary variable married to equal one if a person is married and zero if otherwise.
Table 7.1 gives a partial listing of a wage data set that might result. We see that Person 1 is female and
not married, Person 2 is female and married, Person 3 is male and not married, and so on.
Why do we use the values zero and one to describe qualitative information? In a sense, these
values are arbitrary: any two different values would do. The real benefit of capturing qualitative information using zero-one variables is that it leads to regression models where the parameters have very
natural interpretations, as we will see now.
58860_ch07_hr_220-261.indd 221
10/18/18 4:27 PM
222
PART 1 Regression Analysis with Cross-Sectional Data
Table 7.1 A Partial Listing of the Data in WAGE1
person
wage
educ
exper
female
1
3.10
11
2
1
married
0
2
3.24
12
22
1
1
3
3.00
11
2
0
0
4
6.00
8
44
0
1
5
5.30
12
7
0
1
∙
∙
∙
∙
∙
∙
∙
∙
∙
∙
∙
∙
∙
∙
∙
∙
∙
∙
525
11.56
16
5
0
1
526
3.50
14
5
1
0
7-2 A Single Dummy Independent Variable
How do we incorporate binary information into regression models? In the simplest case, with only
a single dummy explanatory variable, we just add it as an independent variable in the equation. For
example, consider the following simple model of hourly wage determination:
wage 5 b0 1 d0 female 1 b1educ 1 u.
[7.1]
We use d0 as the parameter on female in order to highlight the interpretation of the parameters multiplying dummy variables; later, we will use whatever notation is most convenient.
In model (7.1), only two observed factors affect wage: gender and education. Because female 5 1
when the person is female, and female 5 0 when the person is male, the parameter d0 has the following interpretation: d0 is the difference in hourly wage between females and males, given the same
amount of education (and the same error term u). Thus, the coefficient d0 determines whether there
is discrimination against women: if d0 , 0, then for the same level of other factors, women earn less
than men on average.
In terms of expectations, if we assume the zero conditional mean assumption E 1 u 0 female,
educ 2 5 0, then
d0 5 E 1 wage 0 female 5 1,educ 2 2 E 1 wage 0 female 5 0,educ 2 .
Because female 5 1 corresponds to females and female 5 0 corresponds to males, we can write this
more simply as
d0 5 E 1 wage 0 female,educ 2 2 E 1 wage 0 male,educ 2 .
[7.2]
The key here is that the level of education is the same in both expectations; the difference, d0, is due
to gender only.
The situation can be depicted graphically as an intercept shift between males and females. In
Figure 7.1, the case d0 , 0 is shown, so that men earn a fixed amount more per hour than women. The
difference does not depend on the amount of education, and this explains why the wage-education
profiles for women and men are parallel.
At this point, you may wonder why we do not also include in (7.1) a dummy variable, say male,
which is one for males and zero for females. This would be redundant. In (7.1), the intercept for males
is b0, and the intercept for females is b0 1 d0. Because there are just two groups, we only need two
different intercepts. This means that, in addition to b0, we need to use only one dummy variable; we
have chosen to include the dummy variable for females. Using two dummy variables would introduce
58860_ch07_hr_220-261.indd 222
10/18/18 4:27 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
223
Figure 7.1 Graph of wage 5 b0 1 d0 female 1 b1 educ for d0 , 0.
wage
men: wage =
0
1
educ
1
wage = (
slope =
women:
0 1 0) +
1
educ
1
0
0
1
0
0
educ
perfect collinearity because female 1 male 5 1, which means that male is a perfect linear function of
female. Including dummy variables for both genders is the simplest example of the so-called dummy
variable trap, which arises when too many dummy variables describe a given number of groups. We
will discuss this problem in detail later.
In (7.1), we have chosen males to be the base group or benchmark group, that is, the group
against which comparisons are made. This is why b0 is the intercept for males, and d0 is the difference
in intercepts between females and males. We could choose females as the base group by writing the
model as
wage 5 a0 1 g0 male 1 b1educ 1 u,
where the intercept for females is a0 and the intercept for males is a0 1 g0 ; this implies that
a0 5 b0 1 d0 and a0 1 g0 5 b0. In any application, it does not matter how we choose the base
group, but it is important to keep track of which group is the base group.
Some researchers prefer to drop the overall intercept in the model and to include dummy variables for each group. The equation would then be wage 5 b0male 1 a0female 1 b1educ 1 u, where
the intercept for men is b0 and the intercept for women is a0. There is no dummy variable trap in this
case because we do not have an overall intercept. However, this formulation has little to offer, because
testing for a difference in the intercepts is more difficult, and there is no generally agreed upon way to
compute R‑squared in regressions without an intercept. Therefore, we will always include an overall
intercept for the base group.
Nothing much changes when more explanatory variables are involved. Taking males as the base
group, a model that controls for experience and tenure in addition to education is
wage 5 b0 1 d0 female 1 b1educ 1 b2exper 1 b3 tenure 1 u.
58860_ch07_hr_220-261.indd 223
[7.3]
10/18/18 4:27 PM
224
PART 1 Regression Analysis with Cross-Sectional Data
If educ, exper, and tenure are all relevant productivity characteristics, the null hypothesis of no difference between men and women is H0: d0 5 0. The alternative that there is discrimination against
women is H1: d0 , 0.
How can we actually test for wage discrimination? The answer is simple: just estimate the model
by OLS, exactly as before, and use the usual t statistic. Nothing changes about the mechanics of OLS
or the statistical theory when some of the independent variables are defined as dummy variables. The
only difference with what we have done up until now is in the interpretation of the coefficient on the
dummy variable.
Example 7.1
Hourly Wage Equation
Using the data in WAGE1, we estimate model (7.3). For now, we use wage, rather than log(wage), as
the dependent variable:
wage 5 21.57 2 1.81 female 1 .572 educ 1 0.25 exper 1 .141 tenure
1 .72 2 1 .26 2
1 .049 2
1 .012 2
1 .021 2
[7.4]
2
n 5 526, R 5 .364.
The negative intercept—the intercept for men, in this case—is not very meaningful because no one
has zero values for all of educ, exper, and tenure in the sample. The coefficient on female is interesting because it measures the average difference in hourly wage between a man and a woman who have
the same levels of educ, exper, and tenure. If we take a woman and a man with the same levels of
education, experience, and tenure, the woman earns, on average, $1.81 less per hour than the man.
(Recall that these are 1976 wages.)
It is important to remember that, because we have performed multiple regression and controlled
for educ, exper, and tenure, the $1.81 wage differential cannot be explained by different average levels of education, experience, or tenure between men and women. We can conclude that the differential of $1.81 is due to gender or factors associated with gender that we have not controlled for in the
regression. [In 2013 dollars, the wage differential is about 4.09 1 1.81 2 < 7.40.]
It is informative to compare the coefficient on female in equation (7.4) to the estimate we get
when all other explanatory variables are dropped from the equation:
wage 5 7.10 2 2.51 female
1 .21 2 1 .30 2
[7.5]
2
n 5 526, R 5 .116.
As discussed in Section 2.7, the coefficients in (7.5) have a simple interpretation. The intercept is the
average wage for men in the sample (let female 5 0), so men earn $7.10 per hour on average. The
coefficient on female is the difference in the average wage between women and men. Thus, the average wage for women in the sample is 7.10 2 2.51 5 4.59, or $4.59 per hour. (Incidentally, there are
274 men and 252 women in the sample.)
Equation (7.5) provides a simple way to carry out a comparison-of-means test between the two
groups, which in this case are men and women. The estimated difference, 22.51, has a t statistic of
28.37, which is very statistically significant (and, of course, $2.51 is economically large as well).
Generally, simple regression on a constant and a dummy variable is a straightforward way to compare
the means of two groups. For the usual t test to be valid, we must assume that the homoskedasticity
assumption holds, which means that the population variance in wages for men is the same as that for
women.
The estimated wage differential between men and women is larger in (7.5) than in (7.4) because
(7.5) does not control for differences in education, experience, and tenure, and these are lower, on
average, for women than for men in this sample. Equation (7.4) gives a more reliable estimate of the
ceteris paribus gender wage gap; it still indicates a very large differential.
58860_ch07_hr_220-261.indd 224
10/18/18 4:27 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
225
In many cases, dummy independent variables reflect choices of individuals or other economic
units (as opposed to something predetermined, such as gender). In such situations, the matter of causality is again a central issue. In the following example, we would like to know whether personal
computer ownership causes a higher college grade point average.
Example 7.2
Effects of Computer Ownership on College GPA
In order to determine the effects of computer ownership on college grade point average, we estimate
the model
colGPA 5 b0 1 d0 PC 1 b1hsGPA 1 b2 ACT 1 u,
where the dummy variable PC equals one if a student owns a personal computer and zero otherwise.
There are various reasons PC ownership might have an effect on colGPA. A student’s schoolwork
might be of higher quality if it is done on a computer, and time can be saved by not having to wait at a
computer lab. Of course, a student might be more inclined to play computer games or surf the Internet
if he or she owns a PC, so it is not obvious that d0 is positive. The variables hsGPA (high school GPA)
and ACT (achievement test score) are used as controls: it could be that stronger students, as measured
by high school GPA and ACT scores, are more likely to own computers. We control for these factors
because we would like to know the average effect on colGPA if a student is picked at random and
given a personal computer.
Using the data in GPA1, we obtain
colGPA 5 1.26 1 .157 PC 1 .447 hsGPA 1 .0087 ACT
1 .33 2 1 .057 2
1 .094 2
1 .0105 2
[7.6]
n 5 141, R2 5 .219.
This equation implies that a student who owns a PC has a predicted GPA about .16 points higher than
a comparable student without a PC (remember, both colGPA and hsGPA are on a four-point scale).
The effect is also very statistically significant, with tPC 5 .157/.057 < 2.75.
What happens if we drop hsGPA and ACT from the equation? Clearly, dropping the latter variable should have very little effect, as its coefficient and t statistic are very small. But hsGPA is very
significant, and so dropping it could affect the estimate of bPC. Regressing colGPA on PC gives an
estimate on PC equal to about .170, with a standard error of .063; in this case, b^ PC and its t statistic do
not change by much.
In the exercises at the end of the chapter, you will be asked to control for other factors in the
equation to see if the computer ownership effect disappears, or if it at least gets notably smaller.
Each of the previous examples can be viewed as having relevance for policy analysis. In the
first example, we were interested in gender discrimination in the workforce. In the second example,
we were concerned with the effect of computer ownership on college performance. A special case of
policy analysis is program evaluation, where we would like to know the effect of economic or social
programs on individuals, firms, neighborhoods, cities, and so on.
In the simplest case, there are two groups of subjects. The control group does not participate
in the program. The experimental group or treatment group does take part in the program. These
names come from literature in the experimental sciences, and they should not be taken literally.
Except in rare cases, the choice of the control and treatment groups is not random. However, in some
cases, multiple regression analysis can be used to control for enough other factors in order to estimate
the causal effect of the program.
58860_ch07_hr_220-261.indd 225
10/18/18 4:27 PM
226
PART 1 Regression Analysis with Cross-Sectional Data
Example 7.3
Effects of Training Grants on Hours of Training
Using the 1988 data for Michigan manufacturing firms in JTRAIN, we obtain the following estimated
equation:
hrsemp 5 46.67 1 26.25 grant 2 .98 log 1 sales 2 2 6.07 log 1 employ 2
1 43.41 2 1 5.59 2
1 3.54 2
1 3.88 2
[7.7]
2
n 5 105, R 5 .237.
The dependent variable is hours of training per employee, at the firm level. The variable grant is a
dummy variable equal to one if the firm received a job training grant for 1988, and zero otherwise.
The variables sales and employ represent annual sales and number of employees, respectively. We
cannot enter hrsemp in logarithmic form because hrsemp is zero for 29 of the 105 firms used in the
regression.
The variable grant is very statistically significant, with tgrant 5 4.70. Controlling for sales and
employment, firms that received a grant trained each worker, on average, 26.25 hours more. Because
the average number of hours of per worker training in the sample is about 17, with a maximum value
of 164, grant has a large effect on training, as is expected.
The coefficient on log(sales) is small and very insignificant. The coefficient on log(employ)
means that, if a firm is 10% larger, it trains its workers about .61 hour less. Its t statistic is 21.56,
which is only marginally statistically significant.
As with any other independent variable, we should ask whether the measured effect of a qualitative
variable is causal. In equation (7.7), is the difference in training between firms that receive grants and
those that do not due to the grant, or is grant receipt simply an indicator of something else? It might be
that the firms receiving grants would have, on average, trained their workers more even in the absence
of a grant. Nothing in this analysis tells us whether we have estimated a causal effect; we must know
how the firms receiving grants were determined. We can only hope we have controlled for as many
factors as possible that might be related to whether a firm received a grant and to its levels of training.
In Section 7.6 we return to policy analysis using binary indicators, including obtaining a more flexible framework in the context of potential outcomes. These themes reappear in the remainder of the text.
7-2a Interpreting Coefficients on Dummy Explanatory
Variables When the Dependent Variable Is log(y)
A common specification in applied work has the dependent variable appearing in logarithmic form,
with one or more dummy variables appearing as independent variables. How do we interpret the dummy
variable coefficients in this case? Not surprisingly, the coefficients have a percentage interpretation.
Example 7.4
Housing Price Regression
Using the data in HPRICE1, we obtain the equation
log 1 price 2 5 21.35 1 .168 log 1 lotsize 2 1 .707 log 1 sqrft 2
1 .65 2 1 .038 2
1 .093 2
1 .027 bdrms 1 .054 colonial
1 .029 2
1 .045 2
[7.8]
n 5 88, R2 5 .649.
58860_ch07_hr_220-261.indd 226
10/18/18 4:27 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
227
All the variables are self-explanatory except colonial, which is a binary variable equal to one if the
house is of the colonial style. What does the coefficient on colonial mean? For given levels of lotsize,
sqrft, and bdrms, the difference in log 1 price 2 between a house of colonial style and that of another
style is .054. This means that a colonial-style house is predicted to sell for about 5.4% more, holding
other factors fixed.
This example shows that, when log(y) is the dependent variable in a model, the coefficient on a
dummy variable, when multiplied by 100, is interpreted as the percentage difference in y, holding all
other factors fixed. When the coefficient on a dummy variable suggests a large proportionate change
in y, the exact percentage difference can be obtained exactly as with the semi-elasticity calculation in
Section 6-2.
Example 7.5
Log Hourly Wage Equation
Let us reestimate the wage equation from Example 7.1, using log(wage) as the dependent variable and
adding quadratics in exper and tenure:
log 1 wage 2 5 .417 2 .297 female 1 .080 educ 1 .029 exper
1 .099 2 1 .036 2
1 .007 2
1 .005 2
2 .00058 exper2 1 .032 tenure 2 .00059 tenure2
1 .00010 2
1 .007 2
1 .00023 2
2
n 5 526, R 5 .441.
[7.9]
Using the same approximation as in Example 7.4, the coefficient on female implies that, for the
same levels of educ, exper, and tenure, women earn about 100(.297) 5 29.7% less than men. We
can do better than this by computing the exact percentage difference in predicted wages. What we
want is the proportionate difference in wages between females and males, holding other factors fixed:
1 wageF 2 wageM 2 /wageM. What we have from (7.9) is
log 1 wageF 2 2 log 1 wageM 2 5 2.297.
Exponentiating and subtracting one gives
1 wageF 2 wageM 2 / wageM 5 exp 1 2.297 2 2 1 < 2.257.
This more accurate estimate implies that a woman’s wage is, on average, 25.7% below a comparable
man’s wage.
If we had made the same correction in Example 7.4, we would have obtained exp 1 .054 2 2 1
< .0555, or about 5.6%. The correction has a smaller effect in Example 7.4 than in the wage example
because the magnitude of the coefficient on the dummy variable is much smaller in (7.8) than in (7.9).
Generally, if b^ 1 is the coefficient on a dummy variable, say x1, when log(y) is the dependent variable, the exact percentage difference in the predicted y when x1 5 1 versus when x1 5 0 is
100 # 3 exp 1 b^ 1 2 2 1 4 .
[7.10]
The estimate b^ 1 can be positive or negative, and it is important to preserve its sign in computing (7.10).
The logarithmic approximation has the advantage of providing an estimate between the magnitudes obtained by using each group as the base group. In particular, although equation (7.10)
gives us a better estimate than 100 # b^ 1 of the percentage by which y for x1 5 1 is greater than y for
x1 5 0, (7.10) is not a good estimate if we switch the base group. In Example 7.5, we can estimate
58860_ch07_hr_220-261.indd 227
10/18/18 4:27 PM
228
PART 1 Regression Analysis with Cross-Sectional Data
the percentage by which a man’s wage exceeds a comparable woman’s wage, and this estimate is
100 # 3 exp 1 2b^ 1 2 2 1 4 5 100 # 3 exp 1 .297 2 2 1 4 < 34.6. The approximation, based on 100 # b^ 1, 29.7, is
between 25.7 and 34.6 (and close to the middle). Therefore, it makes sense to report that “the difference in predicted wages between men and women is about 29.7%,” without having to take a stand on
which is the base group.
7-3 Using Dummy Variables for Multiple Categories
We can use several dummy independent variables in the same equation. For example, we could add
the dummy variable married to equation (7.9). The coefficient on married gives the (approximate)
proportional differential in wages between those who are and are not married, holding female, educ,
exper, and tenure fixed. When we estimate this model, the coefficient on married (with standard error
in parentheses) is .053 (.041), and the coefficient on female becomes 2.290 1 .036 2 . Thus, the “marriage premium” is estimated to be about 5.3%, but it is not statistically different from zero 1 t 5 1.29 2 .
An important limitation of this model is that the marriage premium is assumed to be the same for men
and women; this is relaxed in the following example.
Example 7.6
Log Hourly Wage Equation
Let us estimate a model that allows for wage differences among four groups: married men, married
women, single men, and single women. To do this, we must select a base group; we choose single
men. Then, we must define dummy variables for each of the remaining groups. Call these marrmale,
marrfem, and singfem. Putting these three variables into (7.9) (and, of course, dropping female, as it
is now redundant) gives
log 1 wage 2 5 .321 1 .213 marrmale 2 .198 marrfem
1 .100 2 1 .055 2
1 .058 2
2 .110 singfem 1 .079 educ 1 .027 exper 2 .00054 exper2
1 .056 2
1 .007 2
1 .005 2
1 .00011 2 [7.11]
2
1 .029 tenure 2 .00053 tenure
1 .007 2
1 .00023 2
n 5 526, R2 5 .461.
All of the coefficients, with the exception of singfem, have t statistics well above two in absolute
value. The t statistic for singfem is about 21.96, which is just significant at the 5% level against a
two-sided alternative.
To interpret the coefficients on the dummy variables, we must remember that the base group is
single males. Thus, the estimates on the three dummy variables measure the proportionate difference
in wage relative to single males. For example, married men are estimated to earn about 21.3% more
than single men, holding levels of education, experience, and tenure fixed. [The more precise estimate
from (7.10) is about 23.7%.] A married woman, on the other hand, earns a predicted 19.8% less than
a single man with the same levels of the other variables.
Because the base group is represented by the intercept in (7.11), we have included dummy variables for only three of the four groups. If we were to add a dummy variable for single males to (7.11),
we would fall into the dummy variable trap by introducing perfect collinearity. Some regression packages will automatically correct this mistake for you, while others will just tell you there is perfect
collinearity. It is best to carefully specify the dummy variables because then we are forced to properly
interpret the final model.
58860_ch07_hr_220-261.indd 228
10/18/18 4:27 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
229
Even though single men is the base group in (7.11), we can use this equation to obtain the estimated difference between any two groups. Because the overall intercept is common to all groups, we
can ignore that in finding differences. Thus, the estimated proportionate difference between single
and married women is 2.110 2 12.198 2 5 .088, which means that single women earn about 8.8%
more than married women. Unfortunately, we cannot use equation (7.11) for testing whether the estimated difference between single and married women is statistically significant. Knowing the standard
errors on marrfem and singfem is not enough to carry out the test (see Section 4-4). The easiest thing
to do is to choose one of these groups to be the base group and to reestimate the equation. Nothing
substantive changes, but we get the needed estimate and its standard error directly. When we use married women as the base group, we obtain
log 1 wage 2 5 .123 1 .411 marrmale 1 .198 singmale 1 .088 singfem 1 p ,
1 .106 2 1 .056 2
1 .058 2
1 .052 2
where, of course, none of the unreported coefficients or standard errors have changed. The estimate
on singfem is, as expected, .088. Now, we have a standard error to go along with this estimate. The t
statistic for the null that there is no difference in the population between married and single women
is tsingfem 5 .088/.052 < 1.69. This is marginal evidence against the null hypothesis. We also see that
the estimated difference between married men and married women is very statistically significant
1 tmarrmale 5 7.34 2 .
The previous example illustrates a general principle for including dummy variables to indicate different groups: if the regression model is to have different intercepts for, say, g groups or categories, we
need to include g 2 1 dummy variables in the model along with an intercept. The intercept for the base
group is the overall intercept in the model, and the
dummy variable coefficient for a particular group repG o i n g F u rt h e r 7 . 2
resents the estimated difference in intercepts between
that group and the base group. Including g dummy varIn the baseball salary data found in MLB1,
iables along with an intercept will result in the dummy
players are given one of six positions:
variable trap. An alternative is to include g dummy
frstbase, scndbase, thrdbase, shrtstop,
variables and to exclude an overall intercept. Including
outfield, or catcher. To allow for salary differentials across position, with outfieldg dummies without an overall intercept is sometimes
ers as the base group, which dummy
useful, but it has two practical drawbacks. First, it
variables would you include as independent
makes it more cumbersome to test for differences
variables?
relative to a base group. Second, regression packages
usually change the way R-squared is computed when
an overall intercept is not included. In particular, in the
formula R2 5 1 2 SSR/SST, the total sum of squares, SST, is replaced with a total sum of squares that
n
does not center yi about its mean, say, SST0 5 g i51 y2i . The resulting R-squared, say R20 5 1 2 SSR/SST0,
is sometimes called the uncentered R-squared. Unfortunately, R20 is rarely suitable as a goodnessof-fit measure. It is always true that SST0 $ SST with equality only if y 5 0. Often, SST0 is much
larger than SST, which means that R20 is much larger than R2. For example, if in the previous example
we regress log(wage) on marrmale, singmale, marrfem, singfem, and the other explanatory variables—
without an intercept—the reported R-squared from Stata, which is R20, is .948. This high R-squared is
an artifact of not centering the total sum of squares in the calculation. The correct R-squared is given
in equation (7.11) as .461. Some regression packages, including Stata, have an option to force calculation of the centered R-squared even though an overall intercept has not been included, and using
this option is generally a good idea. In the vast majority of cases, any R-squared based on comparing an
SSR and SST should have SST computed by centering the yi about y. We can think of this SST as the
sum of squared residuals obtained if we just use the sample average, y, to predict each yi. Surely we are
58860_ch07_hr_220-261.indd 229
10/18/18 4:27 PM
230
PART 1 Regression Analysis with Cross-Sectional Data
setting the bar pretty low for any model if all we measure is its fit relative to using a constant predictor.
For a model without an intercept that fits poorly, it is possible that SSR . SST, which means R2 would
be negative. The uncentered R-squared will always be between zero and one, which likely explains why
it is usually the default when an intercept is not estimated in regression models.
7-3a Incorporating Ordinal Information by Using Dummy Variables
Suppose that we would like to estimate the effect of city credit ratings on the municipal bond interest
rate (MBR). Several financial companies, such as Moody’s Investors Service and Standard and Poor’s,
rate the quality of debt for local governments, where the ratings depend on things like probability of
default. (Local governments prefer lower interest rates in order to reduce their costs of borrowing.)
For simplicity, suppose that rankings take on the integer values 5 0, 1, 2, 3, 4 6 , with zero being the
worst credit rating and four being the best. This is an example of an ordinal variable. Call this variable CR for concreteness. The question we need to address is: How do we incorporate the variable CR
into a model to explain MBR?
One possibility is to just include CR as we would include any other explanatory variable:
MBR 5 b0 1 b1CR 1 other factors,
where we do not explicitly show what other factors are in the model. Then b1 is the percentage point
change in MBR when CR increases by one unit, holding other factors fixed. Unfortunately, it is rather
hard to interpret a one-unit increase in CR. We know the quantitative meaning of another year of education, or another dollar spent per student, but things like credit ratings typically have only ordinal
meaning. We know that a CR of four is better than a CR of three, but is the difference between four
and three the same as the difference between one and zero? If not, then it might not make sense to
assume that a one-unit increase in CR has a constant effect on MBR.
A better approach, which we can implement because CR takes on relatively few values, is to
define dummy variables for each value of CR. Thus, let CR1 5 1 if CR 5 1, and CR1 5 0 otherwise;
CR2 5 1 if CR 5 2, and CR2 5 0 otherwise; and so on. Effectively, we take the single credit rating
and turn it into five categories. Then, we can estimate the model
MBR 5 b0 1 d1CR1 1 d2CR2 1 d3CR3 1 d4CR4 1 other factors.
[7.12]
Following our rule for including dummy variables in a model, we include four dummy variables
because we have five categories. The omitted category here is a credit rating of zero, and so it is the
base group. (This is why we do not need to define a
dummy variable for this category.) The coefficients
G o i n g F u rt h e r 7 . 3
are easy to interpret: d1 is the difference in MBR
In model (7.12), how would you test the null
(other factors fixed) between a municipality with a
hypothesis that credit rating has no effect on
credit rating of one and a municipality with a credit
MBR?
rating of zero; d2 is the difference in MBR between a
municipality with a credit rating of two and a municipality with a credit rating of zero; and so on. The
movement between each credit rating is allowed to have a different effect, so using (7.12) is much
more flexible than simply putting CR in as a single variable. Once the dummy variables are defined,
estimating (7.12) is straightforward.
Equation (7.12) contains the model with a constant partial effect as a special case. One
way to write the three restrictions that imply a constant partial effect is d2 5 2d1, d3 5 3d1,
and d4 5 4d1 . When we plug these into equation (7.12) and rearrange, we get MBR 5 b0 1
d1 1 CR1 1 2CR2 1 3CR3 1 4CR4 2 1 other factors. Now, the term multiplying d1 is simply the original credit rating variable, CR. To obtain the F statistic for testing the constant partial effect restrictions,
58860_ch07_hr_220-261.indd 230
10/18/18 4:27 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
231
we obtain the unrestricted R-squared from (7.12) and the restricted R-squared from the regression of
MBR on CR and the other factors we have controlled for. The F statistic is obtained as in equation
(4.41) with q 5 3.
Example 7.7
Effects of Physical Attractiveness on Wage
Hamermesh and Biddle (1994) used measures of physical attractiveness in a wage equation. (The
file BEAUTY contains fewer variables but more observations than used by Hamermesh and Biddle.
See Computer Exercise C12.) Each person in the sample was ranked by an interviewer for physical attractiveness, using five categories (homely, quite plain, average, good looking, and strikingly
beautiful or handsome). Because there are so few people at the two extremes, the authors put people
into one of three groups for the regression analysis: average, below average, and above average,
where the base group is average. Using data from the 1977 Quality of Employment Survey, after
controlling for the usual productivity characteristics, Hamermesh and Biddle estimated an equation
for men:
log 1 wage 2 5 b^ 0 2 .164 belavg 1 .016 abvavg 1 other factors
1 .046 2
1 .033 2
n 5 700, R2 5 .403
and an equation for women:
log 1 wage 2 5 b^ 0 2 .124 belavg 1 .035 abvavg 1 other factors
1 .066 2
1 .049 2
n 5 409, R2 5 .330.
The other factors controlled for in the regressions include education, experience, tenure, marital
status, and race; see Table 3 in Hamermesh and Biddle’s paper for a more complete list. In order
to save space, the coefficients on the other variables are not reported in the paper and neither is the
intercept.
For men, those with below average looks are estimated to earn about 16.4% less than an averagelooking man who is the same in other respects (including education, experience, tenure, marital status, and race). The effect is statistically different from zero, with t 5 23.57. Men with above average
looks are estimated to earn only 1.6% more than men with average looks, and the effect is not statistically significant 1 t , .5 2 .
A woman with below average looks earns about 12.4% less than an otherwise comparable
average-looking woman, with t 5 21.88. As was the case for men, the estimate on abvavg is much
smaller in magnitude and not statistically different from zero.
In related work, Biddle and Hamermesh (1998) revisit the effects of looks on earnings using a
more homogeneous group: graduates of a particular law school. The authors continue to find that
physical appearance has an effect on annual earnings, something that is perhaps not too surprising
among people practicing law.
In some cases, the ordinal variable takes on too many values so that a dummy variable cannot be
included for each value. For example, the file LAWSCH85 contains data on median starting salaries
for law school graduates. One of the key explanatory variables is the rank of the law school. Because
each law school has a different rank, we clearly cannot include a dummy variable for each rank. If we
do not wish to put the rank directly in the equation, we can break it down into categories. The following example shows how this is done.
58860_ch07_hr_220-261.indd 231
10/18/18 4:27 PM
232
PART 1 Regression Analysis with Cross-Sectional Data
Example 7.8
Effects of Law School Rankings on Starting Salaries
Define the dummy variables top10, r11_25, r26_40, r41_60, r61_100 to take on the value unity when
the variable rank falls into the appropriate range. We let schools ranked below 100 be the base group.
The estimated equation is
log 1 salary 2 5 9.17 1 .700 top10 1 .594 r11_25 1 .375 r26_40
1 .41 2 1 .053 2
1 .039 2
1 .034 2
1 .263 r41_60 1 .132 r61_100 1 .0057 LSAT
1 .028 2
1 .021 2
1 .0031 2
[7.13]
1 .041 GPA 1 .036 log 1 libvol 2 1 .0008 log 1 cost 2
1 .074 2
1 .026 2
1 .0251 2
2
2
n 5 136, R 5 .911, R 5 .905.
We see immediately that all of the dummy variables defining the different ranks are very statistically significant. The estimate on r61_100 means that, holding LSAT, GPA, libvol, and cost fixed, the
median salary at a law school ranked between 61 and 100 is about 13.2% higher than that at a law
school ranked below 100. The difference between a top 10 school and a below 100 school is quite
large. Using the exact calculation given in equation (7.10) gives exp 1 .700 2 2 1 < 1.014, and so the
predicted median salary is more than 100% higher at a top 10 school than it is at a below 100 school.
As an indication of whether breaking the rank into different groups is an improvement, we can
compare the adjusted R-squared in (7.13) with the adjusted R-squared from including rank as a single
variable: the former is .905 and the latter is .836, so the additional flexibility of (7.13) is warranted.
Interestingly, once the rank is put into the (admittedly somewhat arbitrary) given categories, all of
the other variables become insignificant. In fact, a test for joint significance of LSAT, GPA, log(libvol),
and log(cost) gives a p-value of .055, which is borderline significant. When rank is included in its
original form, the p-value for joint significance is zero to four decimal places.
One final comment about this example: In deriving the properties of ordinary least squares, we
assumed that we had a random sample. The current application violates that assumption because of the
way rank is defined: a school’s rank necessarily depends on the rank of the other schools in the sample,
and so the data cannot represent independent draws from the population of all law schools. This does
not cause any serious problems provided the error term is uncorrelated with the explanatory variables.
7-4 Interactions Involving Dummy Variables
7-4a Interactions among Dummy Variables
Just as variables with quantitative meaning can be interacted in regression models, so can dummy
variables. We have effectively seen an example of this in Example 7.6, where we defined four categories based on marital status and gender. In fact, we can recast that model by adding an interaction
term between female and married to the model where female and married appear separately. This
allows the marriage premium to depend on gender, just as it did in equation (7.11). For purposes of
comparison, the estimated model with the female-married interaction term is
log 1 wage 2 5 .321 2 .110 female 1 .231 married
1 .100 2 1 .056 2
1 .055 2
#
2 .301 female married 1 p ,
[7.14]
1 .072 2
58860_ch07_hr_220-261.indd 232
10/18/18 4:27 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
233
where the rest of the regression is necessarily identical to (7.11). Equation (7.14) shows explicitly
that there is a statistically significant interaction between gender and marital status. This model also
allows us to obtain the estimated wage differential among all four groups, but here we must be careful
to plug in the correct combination of zeros and ones.
Setting female 5 0 and married 5 0 corresponds to the group single men, which is the base
group, as this eliminates female, married, and female . married. We can find the intercept for married
men by setting female 5 0 and married 5 1 in (7.14); this gives an intercept of .321 1 .213 5 .534,
and so on.
Equation (7.14) is just a different way of finding wage differentials across all gender–marital
status combinations. It allows us to easily test the null hypothesis that the gender differential does
not depend on marital status (equivalently, that the marriage differential does not depend on gender).
Equation (7.11) is more convenient for testing for wage differentials between any group and the base
group of single men.
Example 7.9
Effects of Computer Usage on Wages
Krueger (1993) estimates the effects of computer usage on wages. He defines a dummy variable,
which we call compwork, equal to one if an individual uses a computer at work. Another dummy
variable, comphome, equals one if the person uses a computer at home. Using 13,379 people from the
1989 Current Population Survey, Krueger (1993, Table 4) obtains
log 1 wage 2 5 b^ 0 1 .177 compwork 1 .070 comphome
1 .009 2
1 .019 2
1 .017 compwork # comphome 1 other factors.
1 .023 2
[7.15]
(The other factors are the standard ones for wage regressions, including education, experience, gender,
and marital status; see Krueger’s paper for the exact list.) Krueger does not report the intercept because
it is not of any importance; all we need to know is that the base group consists of people who do not
use a computer at home or at work. It is worth noticing that the estimated return to using a computer at
work (but not at home) is about 17.7%. (The more precise estimate is 19.4%.) Similarly, people who
use computers at home but not at work have about a 7% wage premium over those who do not use a
computer at all. The differential between those who use a computer at both places, relative to those
who use a computer in neither place, is about 26.4% (obtained by adding all three coefficients and multiplying by 100), or the more precise estimate 30.2% obtained from equation (7.10).
The interaction term in (7.15) is not statistically significant, nor is it very big economically. But it
is causing little harm by being in the equation.
7-4b Allowing for Different Slopes
We have now seen several examples of how to allow different intercepts for any number of groups
in a multiple regression model. There are also occasions for interacting dummy variables with
explanatory variables that are not dummy variables to allow for difference in slopes. Continuing
with the wage example, suppose that we wish to test whether the return to education is the same for
men and women, allowing for a constant wage differential between men and women (a differential
for which we have already found evidence). For simplicity, we include only education and gender
in the model. What kind of model allows for different returns to education? Consider the model
log 1 wage 2 5 1 b0 1 d0 female 2 1 1 b1 1 d1 female 2 educ 1 u.
58860_ch07_hr_220-261.indd 233
[7.16]
10/18/18 4:27 PM
234
PART 1 Regression Analysis with Cross-Sectional Data
If we plug female 5 0 into (7.16), then we find that the intercept for males is b0, and the slope on
education for males is b1. For females, we plug in female 5 1; thus, the intercept for females is
b0 1 d0, and the slope is b1 1 d1. Therefore, d0 measures the difference in intercepts between women
and men, and d1 measures the difference in the return to education between women and men. Two of
the four cases for the signs of d0 and d1 are presented in Figure 7.2.
Graph (a) shows the case where the intercept for women is below that for men, and the slope of the line
is smaller for women than for men. This means that women earn less than men at all levels of education,
and the gap increases as educ gets larger. In graph (b), the intercept for women is below that for men, but the
slope on education is larger for women. This means that women earn less than men at low levels of education, but the gap narrows as education increases. At some point, a woman earns more than a man with the
same level of education, and this amount of education is easily found once we have the estimated equation.
How can we estimate model (7.16)? To apply OLS, we must write the model with an interaction
between female and educ:
log 1 wage 2 5 b0 1 d0 female 1 b1educ 1 d1 female # educ 1 u.
[7.17]
The parameters can now be estimated from the regression of log(wage) on female, educ, and
female # educ. Obtaining the interaction term is easy in any regression package. Do not be daunted by
the odd nature of female # educ, which is zero for any man in the sample and equal to the level of education for any woman in the sample.
An important hypothesis is that the return to education is the same for women and men. In terms
of model (7.17), this is stated as H0: d1 5 0, which means that the slope of log(wage) with respect to
educ is the same for men and women. Note that this hypothesis puts no restrictions on the difference
in intercepts, d0. A wage differential between men and women is allowed under this null, but it must
be the same at all levels of education. This situation is described by Figure 7.1.
We are also interested in the hypothesis that average wages are identical for men and women who
have the same levels of education. This means that d0 and d1 must both be zero under the null hypothesis. In equation (7.17), we must use an F test to test H0: d0 5 0, d1 5 0. In the model with just an intercept difference, we reject this hypothesis because H0: d0 5 0 is soundly rejected against H1: d0 , 0.
Figure 7.2 Graphs of equation (7.16): (a) d0 , 0, d1 , 0; (b) d0 , 0, d1 . 0.
log(wage)
log(wage)
women
men
men
women
(a)
58860_ch07_hr_220-261.indd 234
educ
(b)
educ
10/18/18 4:27 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
Example 7.10
235
Log Hourly Wage Equation
We add quadratics in experience and tenure to (7.17):
log 1 wage 2 5 .389 2 .227 female 1 .082 educ
1 .119 2 1 .168 2
1 .008 2
2 .0056 female # educ 1 .029 exper 2 .00058 exper2
1 .0131 2
1 .005 2
1 .00011 2
[7.18]
1 .032 tenure 2 .00059 tenure2
1 .007 2
1 .00024 2
n 5 526, R2 5 .441.
The estimated return to education for men in this equation is .082, or 8.2%. For women, it is
.082 2 .0056 5 .0764, or about 7.6%. The difference, 2.56%, or just over one-half a percentage point less for women, is not economically large nor statistically significant: the t statistic is
2.0056/.0131 < 2.43. Thus, we conclude that there is no evidence against the hypothesis that the
return to education is the same for men and women.
The coefficient on female, while remaining economically large, is no longer significant at conventional levels 1 t 5 21.35 2 . Its coefficient and t statistic in the equation without the interaction were
2.297 and 28.25, respectively [see equation (7.9)]. Should we now conclude that there is no statistically significant evidence of lower pay for women at the same levels of educ, exper, and tenure? This
would be a serious error. Because we have added the interaction female∙educ to the equation, the coefficient on female is now estimated much less precisely than it was in equation (7.9): the standard error
has increased by almost fivefold 1 .168/.036 < 4.67 2 . This occurs because female and female # educ
are highly correlated in the sample. In this example, there is a useful way to think about the multicollinearity: in equation (7.17) and the more general equation estimated in (7.18), d0 measures the wage
differential between women and men when educ 5 0. Very few people in the sample have very low
levels of education, so it is not surprising that we have a difficult time estimating the differential at
educ 5 0 (nor is the differential at zero years of education very informative). More interesting would
be to estimate the gender differential at, say, the average education level in the sample (about 12.5).
To do this, we would replace female # educ with female # 1 educ 212.5 2 and rerun the regression; this
only changes the coefficient on female and its standard error. (See Computer Exercise C7.)
If we compute the F statistic for H0: d0 5 0, d1 5 0, we obtain F 5 34.33, which is a huge value
for an F random variable with numerator df 5 2 and denominator df 5 518: the p-value is zero to
four decimal places. In the end, we prefer model (7.9), which allows for a constant wage differential
between women and men.
G o i n g F u rt h e r 7 . 4
How would you augment the model estimated in (7.18) to allow the return to tenure
to differ by gender?
Example 7.11
As a more complicated example involving interactions, we now look at the effects of race and city
racial composition on major league baseball player
salaries.
Effects of Race on Baseball Player Salaries
Using MLB1, the following equation is estimated for the 330 major league baseball players for which
city racial composition statistics are available. The variables black and hispan are binary indicators
for the individual players. (The base group is white players.) The variable percblck is the percentage of the team’s city that is black, and perchisp is the percentage of Hispanics. The other variables
58860_ch07_hr_220-261.indd 235
10/18/18 4:27 PM
236
PART 1 Regression Analysis with Cross-Sectional Data
measure aspects of player productivity and longevity. Here, we are interested in race effects after
controlling for these other factors.
In addition to including black and hispan in the equation, we add the interactions black∙percblck
and hispan∙perchisp. The estimated equation is
log 1 salary 2 5 10.34 1 .0673 years 1 .0089 gamesyr
1 2.18 2 1 .0129 2
1 .0034 2
1 .00095 bavg 1 .0146 hrunsyr 1 .0045 rbisyr
1 .00151 2
1 .0164 2
1 .0076 2
1 .0072 runsyr 1 .0011 fldperc 1 .0075 allstar
1 .0046 2
1 .0021 2
1 .0029 2
2 .198 black 2 .190 hispan 1 .0125 black # percblck
1 .125 2
1 .153 2
1 .0050 2
1 .0201 hispan # perchisp
1 .0098 2
n 5 330, R2 5 .638.
[7.19]
First, we should test whether the four race variables, black, hispan, black∙percblck, and hispan∙perchisp,
are jointly significant. Using the same 330 players, the R-squared when the four race variables are
dropped is .626. Because there are four restrictions and df 5 330 2 13 in the unrestricted model, the
F statistic is about 2.63, which yields a p-value of .034. Thus, these variables are jointly significant at
the 5% level (though not at the 1% level).
How do we interpret the coefficients on the race variables? In the following discussion, all productivity factors are held fixed. First, consider what happens for black players, holding perchisp fixed.
The coefficient 2.198 on black literally means that, if a black player is in a city with no blacks
1 percblck 5 0 2 , then the black player earns about 19.8% less than a comparable white player. As per‑
cblck increases—which means the white population decreases, because perchisp is held fixed—the
salary of blacks increases relative to that for whites. In a city with 10% blacks, log(salary) for blacks
compared to that for whites is 2.198 1 .0125 1 10 2 5 2.073, so salary is about 7.3% less for blacks
than for whites in such a city. When percblck 5 20, blacks earn about 5.2% more than whites. The
largest percentage of blacks in a city is about 74% (Detroit).
Similarly, Hispanics earn less than whites in cities with a low percentage of Hispanics. But we
can easily find the value of perchisp that makes the differential between whites and Hispanics equal
zero: it must make 2.190 1 .0201 perchisp 5 0, which gives perchisp < 9.45. For cities in which
the percentage of Hispanics is less than 9.45%, Hispanics are predicted to earn less than whites (for
a given black population), and the opposite is true if the percentage of Hispanics is above 9.45%.
Twelve of the 22 cities represented in the sample have Hispanic populations that are less than 9.45%
of the total population. The largest percentage of Hispanics is about 31%.
How do we interpret these findings? We cannot simply claim discrimination exists against
blacks and Hispanics, because the estimates imply that whites earn less than blacks and Hispanics
in cities heavily populated by minorities. The importance of city composition on salaries might be
due to player preferences: perhaps the best black players live disproportionately in cities with more
blacks and the best Hispanic players tend to be in cities with more Hispanics. The estimates in (7.19)
allow us to determine that some relationship is present, but we cannot distinguish between these two
hypotheses.
58860_ch07_hr_220-261.indd 236
10/18/18 4:27 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
237
7-4c Testing for Differences in Regression Functions across Groups
The previous examples illustrate that interacting dummy variables with other independent variables
can be a powerful tool. Sometimes, we wish to test the null hypothesis that two populations or groups
follow the same regression function, against the alternative that one or more of the slopes differ across
the groups. We will also see examples of this in Chapter 13, when we discuss pooling different cross
sections over time.
Suppose we want to test whether the same regression model describes college grade point averages for male and female college athletes. The equation is
cumgpa 5 b0 1 b1sat 1 b2hsperc 1 b3tothrs 1 u,
where sat is SAT score, hsperc is high school rank percentile, and tothrs is total hours of college
courses. We know that, to allow for an intercept difference, we can include a dummy variable for
either males or females. If we want any of the slopes to depend on gender, we simply interact the
appropriate variable with, say, female, and include it in the equation.
If we are interested in testing whether there is any difference between men and women, then we
must allow a model where the intercept and all slopes can be different across the two groups:
cumgpa 5 b0 1 d0 female 1 b1sat 1 d1 female # sat 1 b2hsperc
1 d2 female # hsperc 1 b3tothrs 1 d3 female # tothrs 1 u.
[7.20]
The parameter d0 is the difference in the intercept between women and men, d1 is the slope difference
with respect to sat between women and men, and so on. The null hypothesis that cumgpa follows the
same model for males and females is stated as
H0: d0 5 0, d1 5 0, d2 5 0, d3 5 0.
[7.21]
If one of the dj is different from zero, then the model is different for men and women.
Using the spring semester data from the file GPA3, the full model is estimated as
cumgpa 5 1.48 2 .353 female 1 .0011 sat 1 .00075 female # sat
1 0.21 2 1 .411 2
1 .0002 2
1 .00039 2
2.0085 hsperc 2 .00055 female # hsperc 1 .0023 tothrs
1 .0009 2
1 .0014 2
1 .00316 2
[7.22]
2.00012 female # tothrs
1 .00163 2
n 5 366, R2 5 .406, R2 5 .394.
None of the four terms involving the female dummy variable is very statistically significant; only the
female∙sat interaction has a t statistic close to two. But we know better than to rely on the individual
t statistics for testing a joint hypothesis such as (7.21). To compute the F statistic, we must ­estimate the
restricted model, which results from dropping female and all of the interactions; this gives an R2 (the
restricted R2) of about .352, so the F statistic is about 8.14; the p-value is zero to five decimal places,
which causes us to soundly reject (7.21). Thus, men and women athletes do follow ­different GPA
models, even though each term in (7.22) that allows women and men to be different is i­ndividually
insignificant at the 5% level.
58860_ch07_hr_220-261.indd 237
10/18/18 4:27 PM
238
PART 1 Regression Analysis with Cross-Sectional Data
The large standard errors on female and the interaction terms make it difficult to tell exactly
how men and women differ. We must be very careful in interpreting equation (7.22) because, in
obtaining differences between women and men, the interaction terms must be taken into account.
If we look only at the female variable, we would wrongly conclude that cumgpa is about .353
less for women than for men, holding other factors fixed. This is the estimated difference only
when sat, hsperc, and tothrs are all set to zero, which is not close to being a possible scenario. At
sat 5 1,100, hsperc 5 10, and tothrs 5 50, the predicted difference between a woman and a man
is 2.353 1 .00075 1 1,100 2 2 .00055 1 10 2 2 .00012 1 50 2 < .461. That is, the female athlete is predicted to have a GPA that is almost one-half a point higher than the comparable male athlete.
In a model with three variables, sat, hsperc, and tothrs, it is pretty simple to add all of the interactions to test for group differences. In some cases, many more explanatory variables are involved,
and then it is convenient to have a different way to compute the statistic. It turns out that the sum of
squared residuals form of the F statistic can be computed easily even when many independent variables are involved.
In the general model with k explanatory variables and an intercept, suppose we have two groups;
call them g 5 1 and g 5 2. We would like to test whether the intercept and all slopes are the same
across the two groups. Write the model as
y 5 bg, 0 1 bg, 1x1 1 bg, 2x2 1 p 1 bg, kxk 1 u,
[7.23]
for g 5 1 and g 5 2. The hypothesis that each beta in (7.23) is the same across the two groups involves
k 1 1 restrictions (in the GPA example, k 1 1 5 4). The unrestricted model, which we can think of as having a group dummy variable and k interaction terms in addition to the intercept and variables themselves,
has n 2 2 1 k 1 1 2 degrees of freedom. [In the GPA example, n 2 2 1 k 1 1 2 5 366 2 2 1 4 2 5 358.]
So far, there is nothing new. The key insight is that the sum of squared residuals from the unrestricted
model can be obtained from two separate regressions, one for each group. Let SSR1 be the sum of
squared residuals obtained estimating (7.23) for the first group; this involves n1 observations. Let SSR2
be the sum of squared residuals obtained from estimating the model using the second group (n2 observations). In the previous example, if group 1 is females, then n1 5 90 and n2 5 276. Now, the sum of
squared residuals for the unrestricted model is simply SSRur 5 SSR1 1 SSR2. The restricted sum of
squared residuals is just the SSR from pooling the groups and estimating a single equation, say SSRP.
Once we have these, we compute the F statistic as usual:
F5
3 SSRP 2 1 SSR1 1 SSR2 2 4 3 n 2 2 1 k 1 1 2 4
#
,
SSR1 1 SSR2
k11
[7.24]
where n is the total number of observations. This particular F statistic is usually called the Chow
statistic in econometrics. Because the Chow test is just an F test, it is only valid under homoskedasticity. In particular, under the null hypothesis, the error variances for the two groups must be equal. As
usual, normality is not needed for asymptotic analysis.
To apply the Chow statistic to the GPA example, we need the SSR from the regression that pooled
the groups together: this is SSRP 5 85.515. The SSR for the 90 women in the sample is SSR1 5 19.603,
and the SSR for the men is SSR2 5 58.752. Thus, SSRur 5 19.603 1 58.752 5 78.355. The F statistic is 3 1 85.515 2 78.355 2 /78.355 4 1 358/4 2 < 8.18; of course, subject to rounding error, this is what
we get using the R-squared form of the test in the models with and without the interaction terms.
(A word of caution: there is no simple R-squared form of the test if separate regressions have been
estimated for each group; the R-squared form of the test can be used only if interactions have been
included to create the unrestricted model.)
One important limitation of the traditional Chow test, regardless of the method used to implement it, is that the null hypothesis allows for no differences at all between the groups. In many cases,
it is more interesting to allow for an intercept difference between the groups and then to test for slope
58860_ch07_hr_220-261.indd 238
10/18/18 4:27 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
239
differences; we saw one example of this in the wage equation in Example 7.10. There are two ways
to allow the intercepts to differ under the null hypothesis. One is to include the group dummy and all
interaction terms, as in equation (7.22), but then test joint significance of the interaction terms only.
The second approach, which produces an identical statistic, is to form a sum-of-squared-residuals
F statistic, as in equation (7.24), but where the restricted SSR, called “SSRP” in equation (7.24), is
obtained using a regression that contains only an intercept shift. Because we are testing k restrictions,
rather than k 1 1, the F statistic becomes
F5
3 SSRP 2 1 SSR1 1 SSR2 2 4 3 n 2 2 1 k 1 1 2 4
#
.
SSR1 1 SSR2
k
Using this approach in the GPA example, SSRP is obtained from the regression cumgpa on female,
sat, hsperc, and tothrs using the data for both male and female student athletes.
Because there are relatively few explanatory variables in the GPA example, it is easy to estimate
(7.20) and test H0: d1 5 0, d2 5 0, d3 5 0 (with d0 unrestricted under the null). The F statistic for the
three exclusion restrictions gives a p-value equal to .205, and so we do not reject the null hypothesis
at even the 20% significance level.
Failure to reject the hypothesis that the parameters multiplying the interaction terms are all zero
suggests that the best model allows for an intercept difference only:
cumgpa 5 1.39 1 .310 female 1 .0012 sat 2 .0084 hsperc
1 .18 2 1 .059 2
1 .0025 tothrs
1 .0007 2
1 .0002 2
1 .0012 2
[7.25]
n 5 366, R2 5 .398, R2 5 .392.
The slope coefficients in (7.25) are close to those for the base group (males) in (7.22); dropping the
interactions changes very little. However, female in (7.25) is highly significant: its t statistic is over 5,
and the estimate implies that, at given levels of sat, hsperc, and tothrs, a female athlete has a predicted
GPA that is .31 point higher than that of a male athlete. This is a practically important difference.
7-5 A Binary Dependent Variable: The Linear Probability Model
By now, we have learned much about the properties and applicability of the multiple linear regression
model. In the last several sections, we studied how, through the use of binary independent variables,
we can incorporate qualitative information as explanatory variables in a multiple regression model. In
all of the models up until now, the dependent variable y has had quantitative meaning (for example,
y is a dollar amount, a test score, a percentage, or the logs of these). What happens if we want to use
multiple regression to explain a qualitative event?
In the simplest case, and one that often arises in practice, the event we would like to explain is
a binary outcome. In other words, our dependent variable, y, takes on only two values: zero and one.
For example, y can be defined to indicate whether an adult has a high school education; y can indicate
whether a college student used illegal drugs during a given school year; or y can indicate whether a
firm was taken over by another firm during a given year. In each of these examples, we can let y 5 1
denote one of the outcomes and y 5 0 the other outcome.
What does it mean to write down a multiple regression model, such as
y 5 b0 1 b1x1 1 p 1 bkxk 1 u,
58860_ch07_hr_220-261.indd 239
[7.26]
10/18/18 4:27 PM
240
PART 1 Regression Analysis with Cross-Sectional Data
when y is a binary variable? Because y can take on only two values, bj cannot be interpreted as the
change in y given a one-unit increase in xj, holding all other factors fixed: y either changes from zero
to one or from one to zero (or does not change). Nevertheless, the bj still have useful interpretations.
If we assume that the zero conditional mean assumption MLR.4 holds, that is, E 1 u 0 x1, p , xk 2 5 0,
then we have, as always,
E 1 y 0 x 2 5 b0 1 b1x1 1 p 1 bkxk,
where x is shorthand for all of the explanatory variables.
The key point is that when y is a binary variable taking on the values zero and one, it is always
true that P 1 y 5 1 0 x 2 5 E 1 y 0 x 2 : the probability of “success”—that is, the probability that y 5 1—is
the same as the expected value of y. Thus, we have the important equation
P 1 y 5 1 0 x 2 5 b0 1 b1x1 1 p 1 bkxk,
[7.27]
DP 1 y 5 1 0 x 2 5 bjDxj.
[7.28]
which says that the probability of success, say, p 1 x 2 5 P 1 y 5 1 0 x 2 , is a linear function of the xj.
Equation (7.27) is an example of a binary response model, and P 1 y 5 1 0 x 2 is also called the response
probability. (We will cover other binary response models in Chapter 17.) Because probabilities must
sum to one, P 1 y 5 0 0 x 2 5 1 2 P 1 y 5 1 0 x 2 is also a linear function of the xj.
The multiple linear regression model with a binary dependent variable is called the linear probability model (LPM) because the response probability is linear in the parameters bj, In the LPM, bj
measures the change in the probability of success when xj changes, holding other factors fixed:
With this in mind, the multiple regression model can allow us to estimate the effect of various explanatory variables on qualitative events. The mechanics of OLS are the same as before.
If we write the estimated equation as
y^ 5 b^ 0 1 b^ 1x1 1 p 1 b^ kxk,
we must now remember that y^ is the predicted probability of success. Therefore, b^ 0 is the predicted
probability of success when each xj is set to zero, which may or may not be interesting. The slope
coefficient b^ 1 measures the predicted change in the probability of success when x1 increases by
one unit.
To correctly interpret a linear probability model, we must know what constitutes a “success.”
Thus, it is a good idea to give the dependent variable a name that describes the event y 5 1. As an
example, let inlf (“in the labor force”) be a binary variable indicating labor force participation by a
married woman during 1975: inlf 5 1 if the woman reports working for a wage outside the home at
some point during the year, and zero otherwise. We assume that labor force participation depends on
other sources of income, including husband’s earnings (nwifeinc, measured in thousands of dollars),
years of education (educ), past years of labor market experience (exper), age, number of children less
than six years old (kidslt6), and number of kids between 6 and 18 years of age (kidsge6). Using the
data in MROZ from Mroz (1987), we estimate the following linear probability model, where 428 of
the 753 women in the sample report were in the labor force at some point during 1975:
inlf 5 .586 2 .0034 nwifeinc 1 .038 educ 1 .039 exper
1 .154 2 1 .0014 2
1 .007 2
1 .006 2
2 .00060 exper2 2 .016 age 2 .262 kidslt6 1 .013 kidsge6
1 .00018 2
1 .002 2
1 .034 2
1 .013 2
n 5 753, R2 5 .264.
58860_ch07_hr_220-261.indd 240
[7.29]
10/18/18 4:27 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
241
Using the usual t statistics, all variables in (7.29) except kidsge6 are statistically significant, and all
of the significant variables have the effects we would expect based on economic theory (or common
sense).
To interpret the estimates, we must remember that a change in the independent variable changes
the probability that inlf 5 1. For example, the coefficient on educ means that, everything else in
(7.29) held fixed, another year of education increases the probability of labor force participation by
.038. If we take this equation literally, 10 more years of education increases the probability of being
in the labor force by .038 1 10 2 5 .38, which is a pretty large increase in a probability. The relationship between the probability of labor force participation and educ is plotted in Figure 7.3. The other
independent variables are fixed at the values nwifeinc 5 50, exper 5 5, age 5 30, kidslt6 5 1, and
kidsge6 5 0 for illustration purposes. The predicted probability is negative until education equals
3.84 years. This should not cause too much concern because, in this sample, no woman has less
than five years of education. The largest reported education is 17 years, and this leads to a predicted
probability of .5. If we set the other independent variables at different values, the range of predicted
probabilities would change. But the marginal effect of another year of education on the probability of
labor force participation is always .038.
The coefficient on nwifeinc implies that, if Dnwifeinc 5 10 (which means an increase of $10,000),
the probability that a woman is in the labor force falls by .034. This is not an especially large effect
given that an increase in income of $10,000 is substantial in terms of 1975 dollars. Experience has
been entered as a quadratic to allow the effect of past experience to have a diminishing effect on the
labor force participation probability. Holding other factors fixed, the estimated change in the probability is approximated as .039 2 2 1 .0006 2 exper 5 .039 2 .0012 exper. The point at which past experience has no effect on the probability of labor force participation is .039/.0012 5 32.5, which is a high
level of experience: only 13 of the 753 women in the sample have more than 32 years of experience.
Figure 7.3 Estimated relationship between the probability of being in the labor force
and years of education, with other explanatory variables fixed.
probability
of labor
force
participation
.5
slope = .038
0
3.84
educ
–.146
58860_ch07_hr_220-261.indd 241
10/18/18 4:27 PM
242
PART 1 Regression Analysis with Cross-Sectional Data
Unlike the number of older children, the number of young children has a huge impact on labor
force participation. Having one additional child less than six years old reduces the probability of participation by 2.262, at given levels of the other variables. In the sample, just under 20% of the women
have at least one young child.
This example illustrates how easy linear probability models are to estimate and interpret, but
it also highlights some shortcomings of the LPM. First, it is easy to see that, if we plug certain
combinations of values for the independent variables into (7.29), we can get predictions either less
than zero or greater than one. Because these are predicted probabilities, and probabilities must be
between zero and one, this can be a little embarrassing. For example, what would it mean to predict that a woman is in the labor force with a probability of –.10? In fact, of the 753 women in the
sample, 16 of the fitted values from (7.29) are less than zero, and 17 of the fitted values are greater
than one.
A related problem is that a probability cannot be linearly related to the independent variables
for all their possible values. For example, (7.29) predicts that the effect of going from zero children
to one young child reduces the probability of working by .262. This is also the predicted drop if the
woman goes from having one young child to two. It seems more realistic that the first small child
would reduce the probability by a large amount, but subsequent children would have a smaller marginal effect. In fact, when taken to the extreme, (7.29) implies that going from zero to four young
children reduces the probability of working by Dinlf 5 .262 1 Dkidslt6 2 5 .262 1 4 2 5 1.048, which is
impossible.
Even with these problems, the linear probability model is useful and often applied in economics.
It usually works well for values of the independent variables that are near the averages in the sample.
In the labor force participation example, no women in the sample have four young children; in fact,
only three women have three young children. Over 96% of the women have either no young children
or one small child, and so we should probably restrict attention to this case when interpreting the
estimated equation.
Predicted probabilities outside the unit interval are a little troubling when we want
to make predictions. Still, there are ways to use the estimated probabilities (even if some are
negative or greater than one) to predict a zero-one outcome. As before, let y^ i denote the fitted values—which may not be bounded between zero and one. Define a predicted value as
|
y i 5 1 if y^ i $ .5 and |
y i 5 0 if y^ i , .5. Now we have a set of predicted values, |
y i, i 5 1, . . . , n,
that, like the yi, are either zero or one. We can use the data on yI and |
y i to obtain the frequencies
with which we correctly predict yi 5 1 and yi 5 0, as well as the proportion of overall correct
predictions. The latter measure, when turned into a percentage, is a widely used goodness-of-fit
measure for binary dependent variables: the percent correctly predicted. An example is given in
Computer Exercise C9(v), and further discussion, in the context of more advanced models, can be
found in Section 17-1.
Due to the binary nature of y, the linear probability model does violate one of the Gauss-Markov
assumptions. When y is a binary variable, its variance, conditional on x, is
Var 1 y 0 x 2 5 p 1 x 2 3 1 2 p 1 x 2 4 ,
[7.30]
where p(x) is shorthand for the probability of success: p 1 x 2 5 b0 1 b1x1 1 p 1 bkxk. This means
that, except in the case where the probability does not depend on any of the independent variables,
there must be heteroskedasticity in a linear probability model. We know from Chapter 3 that this does
not cause bias in the OLS estimators of the bj. But we also know from Chapters 4 and 5 that homoskedasticity is crucial for justifying the usual t and F statistics, even in large samples. Because the standard errors in (7.29) are not generally valid, we should use them with caution. We will show how to
correct the standard errors for heteroskedasticity in Chapter 8. It turns out that, in many applications,
the usual OLS statistics are not far off, and it is still acceptable in applied work to present a standard
OLS analysis of a linear probability model.
58860_ch07_hr_220-261.indd 242
10/18/18 4:27 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
Example 7.12
243
A Linear Probability Model of Arrests
Let arr86 be a binary variable equal to unity if a man was arrested during 1986, and zero otherwise.
The population is a group of young men in California born in 1960 or 1961 who have at least one
arrest prior to 1986. A linear probability model for describing arr86 is
arr86 5 b0 1 b1 pcnv 1 b2 avgsen 1 b3 tottime 1 b4 ptime86 1 b5 qemp86 1 u,
where
pcnv 5 the proportion of prior arrests that led to a conviction.
avgsen 5 the average sentence served from prior convictions 1 in months 2 .
tottime 5 months spent in prison since age 18 prior to 1986.
ptime86 5 months spent in prison in 1986.
qemp86 5 the number of quarters 1 0 to 4 2 that the man was legally employed in 1986.
The data we use are in CRIME1, the same data set used for Example 3.5. Here, we use a binary
dependent variable because only 7.2% of the men in the sample were arrested more than once. About
27.7% of the men were arrested at least once during 1986. The estimated equation is
arr86 5 .441 2 .162 pcnv 1 .0061 avgsen 2 .0023 tottime
1 .017 2 1 .021 2
1 .0065 2
2 .022 ptime86 2 .043 qemp86
1 .005 2
1 .005 2
1 .0050 2
[7.31]
n 5 2,725, R2 5 .0474.
The intercept, .441, is the predicted probability of arrest for someone who has not been convicted
(and so pcnv and avgsen are both zero), has spent no time in prison since age 18, spent no time in
prison in 1986, and was unemployed during the entire year. The variables avgsen and tottime are
insignificant both individually and jointly (the F test gives p-value 5 .347), and avgsen has a counterintuitive sign if longer sentences are supposed to deter crime. Grogger (1991), using a superset of
these data and different econometric methods, found that tottime has a statistically significant positive
effect on arrests and concluded that tottime is a measure of human capital built up in criminal activity.
Increasing the probability of conviction does lower the probability of arrest, but we must
be careful when interpreting the magnitude of the coefficient. The variable pcnv is a proportion
between zero and one; thus, changing pcnv from zero to one essentially means a change from
no chance of being convicted to being convicted with certainty. Even this large change reduces
the probability of arrest only by .162; increasing pcnv by .5 decreases the probability of arrest
by .081.
The incarcerative effect is given by the coefficient on ptime86. If a man is in prison, he cannot be
arrested. Because ptime86 is measured in months, six more months in prison reduces the probability
of arrest by .022 1 6 2 5 .132. Equation (7.31) gives another example of where the linear probability
model cannot be true over all ranges of the independent variables. If a man is in prison all 12 months
of 1986, he cannot be arrested in 1986. Setting all other variables equal to zero, the predicted probability of arrest when ptime86 5 12 is .441 2 .022 1 12 2 5 .177, which is not zero. Nevertheless, if we
start from the unconditional probability of arrest, .277, 12 months in prison reduces the probability to
essentially zero: .277 2 .022 1 12 2 5 .013.
Finally, employment reduces the probability of arrest in a significant way. All other factors fixed, a
man employed in all four quarters is .172 less likely to be arrested than a man who is not employed at all.
58860_ch07_hr_220-261.indd 243
10/18/18 4:28 PM
244
PART 1 Regression Analysis with Cross-Sectional Data
We can also include dummy independent variables in models with dummy dependent
variables. The coefficient measures the predicted difference in probability relative to the base
group. For example, if we add two race dummies, black and hispan, to the arrest equation, we
obtain
arr86 5 .380 2 .152 pcnv 1 .0046 avgsen 2 .0026 tottime
1 .019 2 1 .021 2
1 .0064 2
1 .0049 2
2 .024 ptime86 2 .038 qemp86 1 .170 black 1 .096 hispan
1 .005 2
1 .005 2
1 .024 2
1 .021 2
[7.32]
n 5 2,725, R2 5 .0682.
G o i n g F u rt h e r 7 . 5
What is the predicted probability of arrest for
a black man with no prior convictions—so
that pcnv, avgsen, tottime, and ptime86 are
all zero—who was employed all four quarters in 1986? Does this seem reasonable?
The coefficient on black means that, all other factors
being equal, a black man has a .17 higher chance of
being arrested than a white man (the base group).
Another way to say this is that the probability of
arrest is 17 percentage points higher for blacks than
for whites. The difference is statistically significant
as well. Similarly, Hispanic men have a .096 higher
chance of being arrested than white men.
7-6 More on Policy Analysis and Program Evaluation
We have seen some examples of models containing dummy variables that can be useful for evaluating
policy. Example 7.3 gave an example of program evaluation, where some firms received job training
grants and others did not.
As we mentioned earlier, we must be careful when evaluating programs because in most examples in the social sciences the control and treatment groups are not randomly assigned. Consider again
the Holzer et al. (1993) study, where we are now interested in the effect of the job training grants on
worker productivity (as opposed to amount of job training). The equation of interest is
log 1 scrap 2 5 b0 1 b1grant 1 b2log 1 sales 2 1 b3log 1 employ 2 1 u,
where scrap is the firm’s scrap rate, and the latter two variables are included as controls. The binary
variable grant indicates whether the firm received a grant in 1988 for job training.
Before we look at the estimates, we might be worried that the unobserved factors affecting
worker productivity—such as average levels of education, ability, experience, and tenure—might be
correlated with whether the firm receives a grant. Holzer et al. point out that grants were given on a
first-come, first-served basis. But this is not the same as giving out grants randomly. It might be that
firms with less productive workers saw an opportunity to improve productivity and therefore were
more diligent in applying for the grants.
Using the data in JTRAIN for 1988—when firms actually were eligible to receive the grants—we
obtain
log 1 scrap 2 5 4.99 2 .052 grant 2 .455 log 1 sales 2
1 4.66 2 1 .431 2
1 .373 2
1 .639 log 1 employ 2
1 .365 2
[7.33]
n 5 50, R2 5 .072.
58860_ch07_hr_220-261.indd 244
10/18/18 4:28 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
245
(Seventeen out of the fifty firms received a training grant, and the average scrap rate is 3.47 across
all firms.) The point estimate of 2.052 on grant means that, for given sales and employ, firms receiving a grant have scrap rates about 5.2% lower than firms without grants. This is the direction of the
expected effect if the training grants are effective, but the t statistic is very small. Thus, from this
cross-sectional analysis, we must conclude that the grants had no effect on firm productivity. We will
return to this example in Chapter 9 and show how adding information from a prior year leads to a
much different conclusion.
Even in cases where the policy analysis does not involve assigning units to a control group and
a treatment group, we must be careful to include factors that might be systematically related to the
binary independent variable of interest. A good example of this is testing for racial discrimination.
Race is something that is not determined by an individual or by government administrators. In fact,
race would appear to be the perfect example of an exogenous explanatory variable, given that it is
determined at birth. However, for historical reasons, race is often related to other relevant factors:
there are systematic differences in backgrounds across race, and these differences can be important in
testing for current discrimination.
As an example, consider testing for discrimination in loan approvals. If we can collect data on,
say, individual mortgage applications, then we can define the dummy dependent variable approved
as equal to one if a mortgage application was approved, and zero otherwise. A systematic difference
in approval rates across races is an indication of discrimination. However, because approval depends
on many other factors, including income, wealth, credit ratings, and a general ability to pay back the
loan, we must control for them if there are systematic differences in these factors across race. A linear
probability model to test for discrimination might look like the following:
approved 5 b0 1 b1nonwhite 1 b2income 1 b3wealth 1 b4credrate 1 other factors.
Discrimination against nonwhites is indicated by a rejection of H0: b1 5 0 in favor of H0: b1 , 0,
because b1 is the amount by which the probability of a nonwhite getting an approval differs from the
probability of a white getting an approval, given the same levels of other variables in the equation. If
income, wealth, and so on are systematically different across races, then it is important to control for
these factors in a multiple regression analysis.
7-6a Program Evaluation and Unrestricted Regression Adjustment
In Section 3-7e, in the context of potential outcomes, we derived an equation that can be used to test
the effectiveness of a policy intervention or a program. Letting w again be the binary policy indicator
and x1, x2, . . . , xk the control variables, we obtained the following population regression function:
E 1 y 0 w, x 2 5 a 1 tw 1 xg 5 a 1 tw 1 g1x1 1 p 1 gkxk,
[7.34]
where y 5 (1 2 w)y(0) 1 wy(1) is the observed outcome and [y(0),y(1)] are the potential or counterfactual outcomes.
The reason for including the xj in (7.34) is to account for the possibility that program participation is not randomly assigned. The problem of participation decisions differing systematically
by individual characteristics is often referred to as the self-selection problem, with “self being
used broadly. For example, children eligible for programs such as Head Start participate largely
based on parental decisions. Because family background and structure play a role in Head Start
participation decisions, and they also tend to predict child outcomes, we should control for these
socioeconomic factors when examining the effects of Head Start [see, for example, Currie and
Thomas (1995)].
In the context of causal inference, the assumption that we have sufficient explanatory variables so that, conditional on those variables, program participation is as good as random, is the
58860_ch07_hr_220-261.indd 245
10/18/18 4:28 PM
246
PART 1 Regression Analysis with Cross-Sectional Data
unconfoundedness or ignorability assumption introduced in Section 3-7e. When we are mainly interested in estimating the effect of a program or intervention, as indicated by w,the explanatory variables
x1, x2, . . . , xk are often called covariates. These are factors that possibly vary with participation decisions as well as with potential outcomes.
The self-selection problem is not restricted to decisions to participate in school or government
programs. It is rampant when studying the economic and societal effects of certain behaviors. For
example, individuals choose to use illegal drugs or to drink alcohol. If we want to examine the effects
of such behaviors on unemployment status, earnings, or criminal behavior, we should be concerned
that drug usage might be correlated with factors affecting potential labor or criminal outcomes.
Without accounting for systematic differences between those who use drugs and those who do not, we
are unlikely to obtain a convincing causal estimate of drug usage.
Self-selection also can be an issue when studying more aggregate units. Cities and states choose
whether to implement certain gun control laws, and it is likely that this decision is systematically
related to other factors that affect violent crime [see, for example, Kleck and Patterson (1993)].
Hospitals choose to be for profit or nonprofit, and this decision may be related to hospital characteristics that affect patient health outcomes.
Most program evaluations are still based on observational (or nonexperimental) data, and so estimating the simple equation
y 5 a 1 tw 1 u
[7.35]
by OLS is unlikely to produce an unbiased or consistent estimate of the causal effect. By including
suitable covariates, estimation of (7.34) is likely to be more convincing. In the context of program
evaluation, using the regression
yi on wi, xi1, xi2, . . . , xik, i 5 1, . . . , n
[7.36]
is a version of what is called regression adjustment, and t^ , the coefficient on wi is the regression
adjusted estimator. The idea is that we have used multiple regression with covariates x1, x2, . . . , xk to
adjust for differences across units in estimating the causal effect.
Recall from Section 3.7e that, in addition to unconfoundedness, equation (7.34) was obtained
under the strong assumption of a constant treatment effect: t = yi(1)2yi (0) for all i. We are now in a
position to relax this assumption. We still maintain the unconfoundedness, or conditional independence, assumption, which we reproduce here for convenience:
w is independent of [y(0), y(1)] conditional on x 5 (x1, . . . , xk)
[7.37]
We also assume the conditional means are linear, but now we allow completely separate equations for
y(0) and y(1). Written in terms of errors u(0) and u(1),
y 1 0 2 5 c0 1 1 x 2 h 2 g0 1 u0 5 c0 1 g0,1 1 x1 2 h1 2 1 p 1 g0,k 1 xk 2 hk 2 1 u 1 0 2 [7.38]
y 1 1 2 5 c1 1 1 x 2 h 2 g1 1 u1 5 c1 1 g1,1 1 x1 2 h1 2 1 p 1 g1,k 1 xk 2 hk 2 1 u 1 1 2 [7.39]
where hj 5 E 1 xj 2 is the population mean of xj, c0 5 E 3 y 1 0 2 4 , and c1 5 E 3 y 1 1 2 4 . The covariates xj
have been centered about their means so that the intercepts, c0 and c1, are the expected values of the
two potential outcomes. Equations (7.38) and (7.39) allow the treatment effect for unit i, yi(1)2yi(0),
to depend on observables xi and the unobservables. For unit i, the treatment effect is
tei 5 yi 1 1 2 2 yi 1 0 2 5 1 c1 2 c0 2 1 1 xi 2 h 2 1 g1 2 g0 2 1 3 ui 1 1 2 2 ui 1 0 2 4
The average treatment effect, which we call t in this section, is t 5 c1 2 c0 because
E 1 tei 2 5 1 c1 2 c0 2 1 E 5 xi 2 h 2 1 g1 2 g0 2 1 3 ui 1 1 2 2 ui 1 0 2 46
5 t 1 0 ? 1 g1 2 g0 2 1 0 5 t,
58860_ch07_hr_220-261.indd 246
10/18/18 4:28 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
247
where 1 xi 2 h 2 has a zero mean by construction and ui(0), ui(1) has zero mean because they are the
errors obtained from conditional expectations. The observed outcome
yi 5 yi(0) 1 wi[yi(1) 2 yi(0)] can be written as
yi 5 c0 1 twi 1 1 xi 2 h 2 g0 1 wi 1 xi 2 h 2 d 1 ui 1 0 2 1 wi 3 ui 1 1 2 2 ui 1 0 2 4 [7.40]
where d 5 (g1 2 g0). If we define
ui 5 ui(0) 1 wi[ui(1) 2 ui(0)]
the unconfoundedness implies
E 1 ui 0 wi,xi 2 5 E 3 ui 1 0 2 0 wi, xi 4 1 wiE 5 3 ui 1 1 2 2 ui 1 0 2 4 0 wi, xi 6 5 E 3 ui 1 0 2 0 xi 4 1 wiE 5 3 ui 1 1 2 2 ui 1 0 2 4 0 xi 6 5 0.
[7.41]
yi on wi, xi1, . . . , xik ,wi ? 1 xi1 2 x1 2 , . . . , wi ? 1 xik 2 xk 2 [7.42]
Equations (7.40) and (7.41) suggest that we run a regression that includes a full set of interactions
between wi and demeaned controls. To implement this method, we also need to replace the unknown
population means, hj, with the sample averages (across the entire sample of n observations), xj. This
leads to the regression
using all n observations. The coefficient on wi, t^ , is the average causal effect or average treatment effect.
We can determine how the treatment effect varies with the xj by multiplying (xj2xj) by the coefficient
on the interaction term, d^ j. Note that we do not have to demean the xj when they appear by themselves,
as failing to do so only changes the intercept in the regression. But it is critical to demean the xj before
constructing the interactions in order to obtain the average treatment effect as the coefficient on wi.
The estimateof t from (7.42) will be different than that from (7.36), as (7.36) omits the k interaction terms. In the literature, the phrase “regression adjustment” often refers to the more flexible
regression in (7.42). For emphasis, one can use the terms restricted regression adjustment (RRA) (and
use the notation t^ rra) and unrestricted regression adjustment (URA) (using t^ ura) for (7.36) and (7.42),
respectively.
It turns out that the estimate t^ from (7.42) can be obtained from two separate regressions, just as
when computing the Chow statistic from Section 7.4c. Working through the details is informative, as
it emphasizes the counterfactual nature of the unrestricted regression adjustment. First, we run regressions separately for the control and treatment groups. For the control group, we use the n0 observations with wi 5 0 and run the regression
yi on xi1, xi2, . . . , xik
and obtain the interercept a^ 0 and k slope estimates g^ 0,1, g^ 0,2, . . . , g^ 0,k. We do the same using the n1
observations with wi = 1, and obtain the intercept a^ 1 and the slopes g^ 1,1, g^ 1,2, . . . ,g^ 1,k.
Now here is where we use counterfactual reasoning: for every unit i in the sample, we predict
yi(0) and yi(1) regardless of whether the unit was in the control or treatment group. Define the predicted values as
^ 0,k xik
y^ i102 5 a^ 0 1 g^ 0,1xi1 1 g^ 0,2xi2 1 p 1 g
112
p
y^ i 5 a^1 1 g^ 1,1xi1 1 g^ 1,2xi2 1 1 g^ 1,k xik
for all i. In other words, we plug the explanatory variables for unit i into both regression functions to
predict the outcomes in the two states of the world: the control state and the treated state. It is then
natural to estimate the average treatment effect as
n
n 21 a 3 y^ i1122y^ i102 4 5 1 a^ 1 2 a^ 0 2 1 1 g^ 1,1 2 g^ 0,1 2 x1 1 p 1 1 g^ 1,k 2 g^ 0,k 2 xk.
i51
58860_ch07_hr_220-261.indd 247
[7.43]
10/18/18 4:28 PM
248
PART 1 Regression Analysis with Cross-Sectional Data
With some algebra, one can show that (7.43) produces the t^ from regression (7.40). Thus, two seemingly different ways of using multiple regression to adjust for differences in units lead to t­he same
estimate of the ATE.
Most regression packages are designed to make calculation of (7.43) easy, because they compute
a predicted value for all observations that have information on the {xij: j 5 1, . . . , k} whether or not
unit i was used in the estimation. However, obtaining a proper standard error when computing (7.43)
“by hand” can be tricky, although some econometrics software has the calculation built in. Regression
(7.42) always produces a valid standard error of t^ but requires obtaining the k interaction terms after
demeaning each xj. Incidentally, we can obtain the Chow test that allows different intercepts by testing whether the k interaction terms are jointly significant using an F statistic. If we fail to reject, we
could return to the regression (7.36) that imposes common slopes.
EXAMPLE 7.13
Evaluating a Job Training Program using Unrestricted Regression Adjustment
The data in JTRAIN98 were used in Examples 3.11 and 4.11 to estimate the effects of a job training program. The variable we would like to explain, y 5 earn98, is labor market earnings in 1998,
the year following the job training program (which took place in 1997). The earnings variable is
measured in thousands of dollars. The variable w = train is the binary participation (or “treatment”)
indicator. We use the same controls as in Example 4.11—earn96, educ, age, and married—but now
we use unrestricted regression adjustment. For comparison purposes, the simple difference-in-means
estimate is tdiffmeams 5 22.05 (se 5 0.48) and the restricted regression adjusted estimate, reported in
equation (4.52), is t^ rra 5 2.44 (se 5 0.44). The estimated equation with full interactions is
earn98 5 5.08 1 3.11 train 1 .353 earn96 1 .378 educ 2 .196 age 1 2.76 married
(1.39) (0.53) (.020) (.078) (.023) (0.55)
1 .133 train ? 1 earn96 2 earn96 2 2 .035 train ? 1 educ 2 educ 2 (0.054)
(.137)
[7.44]
1 .058 train ? 1 age 2 age 2 2 .993 train ? 1 married 2 married 2
(.041)
(.883)
n 5 1, 130, R2 5 0.409.
The estimated average treatment effect is the coefficient on train, t^ ura 5 3.11 (se 5 0.53), which is
very statistically significant with ttrain . 5.8. It is also notably higher than the restricted RA estimate,
although the F statistic for joint significance of the interaction terms gives p-value ≈ 0.113, and so the
interaction terms are not jointly significant at the 10% level.
Example 7.13 warrants some final comments. First, to obtain the average treatment effect as
the coefficient on train, all explanatory variables, including the dummy variable married, must be
demeaned before creating the interaction term with train. Using train ? married in place of the final
interaction forces the coefficient on train to be the average treatment effect for unmarried men—
where the averages are across educ96, educ, and age. The estimate turns out to be 3.79 (se 5 0.81).
The coefficient on train ? married would be unchanged from that in (7.44), 2. 993 (se 5 0.883), and
is still interpreted as the difference in the ATE between married and unmarried men.
In using regression adjustment to estimate the effects of something like a job training program,
there is always the possibility that our control variables, x1, x2, . . . , xk, are not sufficient for fully
overcoming self-selection into participation. One must be on guard at all times unless w is known to
have been randomized. With observational data, the possibilityof finding a spurious effect—in either
58860_ch07_hr_220-261.indd 248
10/18/18 4:28 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
249
direction—is often quite high, even with a rich set of xj. A good example of this is contained in
Currie and Cole (1993). These authors examine the effect of AFDC (Aid to Families with Dependent
Children) participation on the birth weight of a child. Even after controlling for a variety of family
and background characteristics, the authors obtain OLS estimates that imply participation in AFDC
lowers birth weight. As the authors point out, it is hard to believe that AFDC participation itself causes
lower birth weight. [See Currie (1995) for additional examples.]
Using a different econometric method that we discuss in Chapter 15, Currie and Cole find evidence for either no effect or a positive effect of AFDC participation on birth weight. When the selfselection problem causes standard multiple regression analysis to be biased due to a lack of sufficient
control variables, the more advanced methods covered in Chapters 13,14, and 15, can be used instead.
7-7 Interpreting Regression Results with Discrete Dependent Variables
A binary response is the most extreme form of a discrete random variable: it takes on only two values, zero and one. As we discussed in Section 7-5, the parameters in a linear probability model can
be interpreted as measuring the change in the probability that y 5 1 due to a one-unit increase in an
explanatory variable. We also discussed that, because y is a zero-one outcome, P 1 y 5 1 2 5 E 1 y 2 , and
this equality continues to hold when we condition on explanatory variables.
Other discrete dependent variables arise in practice, and we have already seen some examples,
such as the number of times someone is arrested in a given year (Example 3.5). Studies on factors
affecting fertility often use the number of living children as the dependent variable in a regression
analysis. As with number of arrests, the number of living children takes on a small set of integer values, and zero is a common value. The data in FERTIL2, which contains information on a large sample
of women in Botswana is one such example. Often demographers are interested in the effects of education on fertility, with special attention to trying to determine whether education has a causal effect
on fertility. Such examples raise a question about how one interprets regression coefficients: after all,
one cannot have a fraction of a child.
To illustrate the issues, the regression below uses the data in FERTIL2:
children 5 21.997 1 .175 age 2 .090 educ
1 .094 2 1 .003 2
1 .006 2
[7.45]
2
n 5 4,361, R 5 .560.
At this time, we ignore the issue of whether this regression adequately controls for all factors that
affect fertility. Instead we focus on interpreting the regression coefficients.
Consider the main coefficient of interest, b^ educ 5 2.090. If we take this estimate literally, it says
that each additional year of education reduces the estimated number of children by .090—something
obviously impossible for any particular woman. A similar problem arises when trying to interpret
b^ age 5 .175. How can we make sense of these coefficients?
To interpret regression results generally, even in cases where y is discrete and takes on a small
number of values, it is useful to remember the interpretation of OLS as estimating the effects of the xj
on the expected (or average) value of y. Generally, under Assumptions MLR.1 and MLR.4,
E 1 y 0 x1, x2, p , xk 2 5 b0 1 b1x1 1 p 1 bkxk.
[7.46]
Therefore, bj is the effect of a ceteris paribus increase of xj on the expected value of y. As we discussed
in Section 6-4, for a given set of xj values we interpret the predicted value, b^ 0 1 b^ 1x1 1 p 1 b^ kxk, as
an estimate of E 1 y 0 x1, x2, . . . , xk 2 . Therefore, b^ j is our estimate of how the average of y changes when
Dxj 5 1 (keeping other factors fixed).
58860_ch07_hr_220-261.indd 249
10/18/18 4:28 PM
250
PART 1 Regression Analysis with Cross-Sectional Data
Seen in this light, we can now provide meaning to regression results as in equation (7.45). The
coefficient b^ educ 5 2.090 means that we estimate that average fertility falls by .09 children given
one more year of education. A nice way to summarize this interpretation is that if each woman in
a group of 100 obtains another year of education, we estimate there will be nine fewer children
among them.
Adding dummy variables to regressions when y is itself discrete causes no problems when we
interpret the estimated effect in terms of average values. Using the data in FERTIL2 we get
children 5 22.071 1 .177 age 2 .079 educ 2 .362 electric
1 .095 2 1 .003 2
1 .006 2
1 .068 2
[7.47]
n 5 4,358, R2 5 .562,
where electric is a dummy variable equal to one if the woman lives in a home with electricity. Of
course it cannot be true that a particular woman who has electricity has .362 less children than an
otherwise comparable woman who does not. But we can say that when comparing 100 women with
electricity to 100 women without—at the same age and level of education—we estimate the former
group to have about 36 fewer children.
Incidentally, when y is discrete the linear model does not always provide the best estimates of partial effects on E 1 y 0 x1, x2, p , xk 2 . Chapter 17 contains more advanced models and estimation methods
that tend to fit the data better when the range of y is limited in some substantive way. Nevertheless, a
linear model estimated by OLS often provides a good approximation to the true partial effects, at least
on average.
Summary
In this chapter, we have learned how to use qualitative information in regression analysis. In the simplest case, a dummy variable is defined to distinguish between two groups, and the coefficient estimate
on the dummy variable estimates the ceteris paribus difference between the two groups. Allowing for
more than two groups is accomplished by defining a set of dummy variables: if there are g groups, then
g 2 1 dummy variables are included in the model. All estimates on the dummy variables are interpreted relative to the base or benchmark group (the group for which no dummy variable is included in
the model).
Dummy variables are also useful for incorporating ordinal information, such as a credit or a beauty
rating, in regression models. We simply define a set of dummy variables representing different outcomes of
the ordinal variable, allowing one of the categories to be the base group.
Dummy variables can be interacted with quantitative variables to allow slope differences across
different groups. In the extreme case, we can allow each group to have its own slope on every variable, as well as its own intercept. The Chow test can be used to detect whether there are any differences across groups. In many cases, it is more interesting to test whether, after allowing for an
intercept difference, the slopes for two different groups are the same. A standard F test can be used
for this purpose in an unrestricted model that includes interactions between the group dummy and all
variables.
The linear probability model, which is simply estimated by OLS, allows us to explain a binary response using regression analysis. The OLS estimates are now interpreted as changes in the probability of
“success” 1 y 5 1 2 , given a one-unit increase in the corresponding explanatory variable. The LPM does
have some drawbacks: it can produce predicted probabilities that are less than zero or greater than one,
it implies a constant marginal effect of each explanatory variable that appears in its original form, and it
contains heteroskedasticity. The first two problems are often not serious when we are obtaining estimates
58860_ch07_hr_220-261.indd 250
10/18/18 4:28 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
251
of the partial effects of the explanatory variables for the middle ranges of the data. Heteroskedasticity does
invalidate the usual OLS standard errors and test statistics, but, as we will see in the next chapter, this is
easily fixed in large enough samples.
Section 7.6 provides a discussion of how binary variables are used to evaluate policies and programs.
As in all regression analysis, we must remember that program participation, or some other binary regressor
with policy implications, might be correlated with unobserved factors that affect the dependent variable,
resulting in the usual omitted variables bias.
We ended this chapter with a general discussion of how to interpret regression equations when the
dependent variable is discrete. The key is to remember that the coefficients can be interpreted as the effects
on the expected value of the dependent variable.
Key Terms
Base Group
Benchmark Group
Binary Variable
Chow Statistic
Control Group
Covariates
Difference in Slopes
Dummy Variable Trap
Dummy Variables
Experimental Group
Interaction Term
Intercept Shift
Linear Probability Model (LPM)
Ordinal Variable
Percent Correctly Predicted
Policy Analysis
Program Evaluation
Regression adjustment
Response Probability
Self-Selection Problem
Treatment Group
Uncentered R-Squared
Zero-One Variable
Problems
1 Using the data in SLEEP75 (see also Problem 3 in Chapter 3), we obtain the estimated equation
sleep 5 3,840.83 2 .163 totwrk 2 11.71 educ 2 8.70 age
1 235.11 2 1 .018 2
1 5.86 2
1 11.21 2
2
1 .128 age 1 87.75 male
1 .134 2
1 34.33 2
n 5 706, R2 5 .123, R2 5 .117.
The variable sleep is total minutes per week spent sleeping at night, totwrk is total weekly minutes
spent working, educ and age are measured in years, and male is a gender dummy.
(i)
All other factors being equal, is there evidence that men sleep more than women? How strong
is the evidence?
(ii) Is there a statistically significant tradeoff between working and sleeping? What is the estimated
tradeoff?
(iii) What other regression do you need to run to test the null hypothesis that, holding other factors
fixed, age has no effect on sleeping?
2 The following equations were estimated using the data in BWGHT:
log 1 bwght 2 5 4.66 2 .0044 cigs 1 .0093 log 1 faminc 2 1 .016 parity
1 .22 2 1 .0009 2
1 .0059 2
1 .006 2
1 .027 male 1 .055 white
1 .010 2
1 .013 2
n 5 1,388, R2 5 .0472
58860_ch07_hr_220-261.indd 251
10/18/18 4:28 PM
252
PART 1 Regression Analysis with Cross-Sectional Data
and
log 1 bwght 2 5 4.65 2 .0052 cigs 1 .0110 log 1 faminc 2 1 .017 parity
1 .38 2 1 .0010 2
1 .0085 2
1 .006 2
1 .034 male 1 .045 white 2 .0030 motheduc 1 .0032 fatheduc
1 .011 2
2
1 .015 2
n 5 1,191, R 5 .0493.
1 .0030 2
1 .0026 2
The variables are defined as in Example 4.9, but we have added a dummy variable for whether the
child is male and a dummy variable indicating whether the child is classified as white.
(i) In the first equation, interpret the coefficient on the variable cigs. In particular, what is the effect
on birth weight from smoking 10 more cigarettes per day?
(ii) How much more is a white child predicted to weigh than a nonwhite child, holding the other
factors in the first equation fixed? Is the difference statistically significant?
(iii) Comment on the estimated effect and statistical significance of motheduc.
(iv) From the given information, why are you unable to compute the F statistic for joint significance
of motheduc and fatheduc? What would you have to do to compute the F statistic?
3 Using the data in GPA2, the following equation was estimated:
sat 5 1,028.10 1 19.30 hsize 2 2.19 hsize2 2 45.09 female
1 6.29 2
1 3.83 2
1 .53 2
2 169.81 black 1 62.31 female # black
1 12.71 2
2
1 18.15 2
1 4.29 2
n 5 4,137, R 5 .0858.
The variable sat is the combined SAT score; hsize is size of the student’s high school graduating class,
in hundreds; female is a gender dummy variable; and black is a race dummy variable equal to one for
blacks, and zero otherwise.
(i) Is there strong evidence that hsize2 should be included in the model? From this equation, what is
the optimal high school size?
(ii) Holding hsize fixed, what is the estimated difference in SAT score between nonblack females
and nonblack males? How statistically significant is this estimated difference?
(iii) What is the estimated difference in SAT score between nonblack males and black males? Test
the null hypothesis that there is no difference between their scores, against the alternative that
there is a difference.
(iv) What is the estimated difference in SAT score between black females and nonblack females?
What would you need to do to test whether the difference is statistically significant?
4 An equation explaining chief executive officer salary is
log 1 salary 2 5 4.59 1 .257 log 1 sales 2 1 .011 roe 1 .158 finance
1 .30 2 1 .032 2
1 .004 2
1 .181 consprod 2 .283 utility
1 .085 2
n 5 209, R2 5 .357.
1 .099 2
1 .089 2
The data used are in CEOSAL1, where finance, consprod, and utility are binary variables indicating
the financial, consumer products, and utilities industries. The omitted industry is transportation.
58860_ch07_hr_220-261.indd 252
10/18/18 4:28 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
253
(i)
Compute the approximate percentage difference in estimated salary between the utility and
transportation industries, holding sales and roe fixed. Is the difference statistically significant at
the 1% level?
(ii) Use equation (7.10) to obtain the exact percentage difference in estimated salary between the
utility and transportation industries and compare this with the answer obtained in part (i).
(iii) What is the approximate percentage difference in estimated salary between the consumer
products and finance industries? Write an equation that would allow you to test whether the
difference is statistically significant.
5 In Example 7.2, let noPC be a dummy variable equal to one if the student does not own a PC, and zero
otherwise.
(i) If noPC is used in place of PC in equation (7.6), what happens to the intercept in the estimated
equation? What will be the coefficient on noPC? (Hint: Write PC 5 1 2 noPC and plug this
into the equation colGPA 5 b^ 0 1 d^ 0PC 1 b^ 1hsGPA 1 b^ 2ACT.)
(ii) What will happen to the R-squared if noPC is used in place of PC?
(iii) Should PC and noPC both be included as independent variables in the model? Explain.
6 To test the effectiveness of a job training program on the subsequent wages of workers, we specify the
model
log 1 wage 2 5 b0 1 b1train 1 b2educ 1 b3exper 1 u,
where train is a binary variable equal to unity if a worker participated in the program. Think of the
error term u as containing unobserved worker ability. If less able workers have a greater chance of
being selected for the program, and you use an OLS analysis, what can you say about the likely bias in
the OLS estimator of b1? (Hint: Refer back to Chapter 3.)
7 In the example in equation (7.29), suppose that we define outlf to be one if the woman is out of the
labor force, and zero otherwise.
(i) If we regress outlf on all of the independent variables in equation (7.29), what will happen to the
intercept and slope estimates? (Hint: inlf 5 1 2 outlf . Plug this into the population equation
inlf 5 b0 1 b1nwifeinc 1 b2educ 1 … and rearrange.)
(ii) What will happen to the standard errors on the intercept and slope estimates?
(iii) What will happen to the R-squared?
8 Suppose you collect data from a survey on wages, education, experience, and gender. In addition, you
ask for information about marijuana usage. The original question is: “On how many separate occasions
last month did you smoke marijuana?”
(i) Write an equation that would allow you to estimate the effects of marijuana usage on wage,
while controlling for other factors. You should be able to make statements such as, “Smoking
marijuana five more times per month is estimated to change wage by x%.”
(ii) Write a model that would allow you to test whether drug usage has different effects on wages
for men and women. How would you test that there are no differences in the effects of drug
usage for men and women?
(iii) Suppose you think it is better to measure marijuana usage by putting people into one of four
categories: nonuser, light user (1 to 5 times per month), moderate user (6 to 10 times per
month), and heavy user (more than 10 times per month). Now, write a model that allows you to
estimate the effects of marijuana usage on wage.
(iv) Using the model in part (iii), explain in detail how to test the null hypothesis that marijuana
usage has no effect on wage. Be very specific and include a careful listing of degrees of
freedom.
(v) What are some potential problems with drawing causal inference using the survey data that you
collected?
58860_ch07_hr_220-261.indd 253
10/18/18 4:28 PM
254
PART 1 Regression Analysis with Cross-Sectional Data
9 Let d be a dummy (binary) variable and let z be a quantitative variable. Consider the model
y 5 b0 1 d0d 1 b1z 1 d1d # z 1 u;
this is a general version of a model with an interaction between a dummy variable and a quantitative
variable. [An example is in equation (7.17).]
(i) Because it changes nothing important, set the error to zero, u 5 0. Then, when d 5 0 we
can write the relationship between y and z as the function f0 1 z 2 5 b0 1 b1z. Write the same
relationship when d 5 1, where you should use f1 1 z 2 on the left-hand side to denote the linear
function of z.
(ii) Assuming that d1 2 0 (which means the two lines are not parallel), show that the value of z*
such that f0 1 z* 2 5 f1 1 z* 2 is z* 5 2d0/d1. This is the point at which the two lines intersect [as
in Figure 7.2 (b)]. Argue that z* is positive if and only if d0 and d1 have opposite signs.
(iii) Using the data in TWOYEAR, the following equation can be estimated:
log 1 wage 2 5 2.289 2 .357 female 1 .50 totcoll 1 .030 female # totcoll
1 0.011 2 1 .015 2
2
n 5 6,763, R 5 .202,
(iv)
1 .003 2
1 .005 2
where all coefficients and standard errors have been rounded to three decimal places. Using this
equation, find the value of totcoll such that the predicted values of log(wage) are the same for
men and women.
Based on the equation in part (iii), can women realistically get enough years of college so that
their earnings catch up to those of men? Explain.
10 For a child i living in a particular school district, let voucheri be a dummy variable equal to one if a
child is selected to participate in a school voucher program, and let scorei be that child’s score on a subsequent standardized exam. Suppose that the participation variable, voucheri, is completely randomized
in the sense that it is independent of both observed and unobserved factors that can affect the test score.
(i) If you run a simple regression scorei on voucheri using a random sample of size n, does the OLS
estimator provide an unbiased estimator of the effect of the voucher program?
(ii) Suppose you can collect additional background information, such as family income, family
structure (e.g., whether the child lives with both parents), and parents’ education levels. Do you
need to control for these factors to obtain an unbiased estimator of the effects of the voucher
program? Explain.
(iii) Why should you include the family background variables in the regression? Is there a situation
in which you would not include the background variables?
11 The following equations were estimated using the data in ECONMATH, with standard errors r­ eported
under coefficients. The average class score, measured as a percentage, is about 72.2; exactly 50% of
the students are male; and the average of colgpa (grade point average at the start of the term) is about
2.81.
score 5 32.31 1 14.32 colgpa
1 2.00 2
1 0.70 2
n 5 856, R2 5 .329, R2 5 .328.
score 5 29.66 1 3.83 male 1 14.57 colgpa
1 2.04 2 1 0.74 2
1 0.69 2
n 5 856, R2 5 .349, R2 5 .348.
58860_ch07_hr_220-261.indd 254
10/18/18 4:28 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
255
score 5 30.36 1 2.47 male 1 14.33 colgpa 1 0.479 male # colgpa
1 2.86 2 1 3.96 2
1 0.98 2
1 1.383 2
n 5 856, R2 5 .349, R2 5 .347.
score 5 30.36 1 3.82 male 1 14.33 colgpa 1 0.479 male # 1 colgpa 2 2.81 2
1 2.86 2 1 0.74 2
1 0.98 2
1 1.383 2
n 5 856, R2 5 .349, R2 5 .347.
(i)
Interpret the coefficient on male in the second equation and construct a 95% confidence interval
for bmale. Does the confidence interval exclude zero?
(ii) In the second equation, why is the estimate on male so imprecise? Should we now conclude
that there are no gender differences in score after controlling for colgpa? [Hint: You might want
to compute an F statistic for the null hypothesis that there is no gender difference in the model
with the interaction.]
(iii) Compared with the third equation, why is the coefficient on male in the last equation so much
closer to that in the second equation and just as precisely estimated?
12 Consider Example 7.11, where, prior to computing the interaction between the race/ethnicity of a
player and the city’s racial composition, we center the city composition variables about the sample averages, percblck and perchisp (which are, approximately, 16.55 and 10.82, respectively). The resulting
estimated equation is
log(salary) 5 10.23 1 .0673 years 1 .0089 gamesyr 1 .00095 bavg 1 .0146 hrunsyr
(2.18) (.0129) (.0034) (.00151)
(.0164)
1 .0045 rbisyr 1 .0072 runsyr 1 .0011 fldperc 1 .0075 allstar
(.0076) (0.0046) (.0021) (.0029)
1 .0080 black 1 .0273 hispan 1 .0125 black ? 1 percblck 2 percblck 2
(.0840) (.1084) (.0050)
1 .0201 hispan ? 1 perchisp 2 perchisp 2
(.0098)
n 5 330, R2 5 0.638.
(i)
Why are the coefficients on black and hispan now so much different than those reported in
equation (7.19)? In particular, how can you interpret these coefficients?
(ii) What do you make of the fact that neither black nor hispan is statistically significant in the
above equation?
(iii) In comparing the above equation to (7.19), has anything else changed? Why or why not?
13 (i)
In the context of potential outcomes with a sample of size n, let [yi(0), yi(1)] denote the pair of
potential outcomes for unit i. Define the averages
y 1 0 2 5 n 21 a yi 1 0 2
n
y112 5 n
21
i51
n
a yi 1 1 2
i51
and define the sample average treatment effect (SATE) as SATE = y 1 1 2 – y 1 0 2 . Can you compute
the SATE given the typical program evaluation data set?
(ii)
58860_ch07_hr_220-261.indd 255
Let y0 and y1 be the sample averages of the observed yi for the control and treated groups,
respectively. Show howthese differ from y 1 0 2 and y 1 1 2 .
10/18/18 4:28 PM
256
PART 1 Regression Analysis with Cross-Sectional Data
Computer Exercises
C1 Use the data in GPA1 for this exercise.
(i) Add the variables mothcoll and fathcoll to the equation estimated in (7.6) and report the results
in the usual form. What happens to the estimated effect of PC ownership? Is PC still statistically
significant?
(ii) Test for joint significance of mothcoll and fathcoll in the equation from part (i) and be sure to
report the p-value.
(iii) Add hsGPA2 to the model from part (i) and decide whether this generalization is needed.
C2 Use the data in WAGE2 for this exercise.
(i) Estimate the model
log 1 wage 2 5 b0 1 b1educ 1 b2exper 1 b3tenure 1 b4married
1 b5black 1 b6south 1 b7urban 1 u
and report the results in the usual form. Holding other factors fixed, what is the approximate
difference in monthly salary between blacks and nonblacks? Is this difference statistically
significant?
(ii) Add the variables exper2 and tenure2 to the equation and show that they are jointly insignificant
at even the 20% level.
(iii) Extend the original model to allow the return to education to depend on race and test whether
the return to education does depend on race.
(iv) Again, start with the original model, but now allow wages to differ across four groups of people:
married and black, married and nonblack, single and black, and single and nonblack. What is
the estimated wage differential between married blacks and married nonblacks?
C3 A model that allows major league baseball player salary to differ by position is
log 1 salary 2 5 b0 1 b1 years 1 b2 gamesyr 1 b3 bavg 1 b4 hrunsyr
1 b5 rbisyr 1 b6 runsyr 1 b7 fldperc 1 b8 allstar
1 b9 frstbase 1 b10scndbase 1 b11thrdbase 1 b12 shrtstop
1 b13 catcher 1 u,
where outfield is the base group.
(i) State the null hypothesis that, controlling for other factors, catchers and outfielders earn, on
average, the same amount. Test this hypothesis using the data in MLB1 and comment on the
size of the estimated salary differential.
(ii) State and test the null hypothesis that there is no difference in average salary across positions,
once other factors have been controlled for.
(iii) Are the results from parts (i) and (ii) consistent? If not, explain what is happening.
C4 Use the data in GPA2 for this exercise.
(i) Consider the equation
colgpa 5 b0 1 b1hsize 1 b2hsize2 1 b3hsperc 1 b4sat
1 b5female 1 b6athlete 1 u,
where colgpa is cumulative college grade point average; hsize is size of high school graduating
class, in hundreds; hsperc is academic percentile in graduating class; sat is combined SAT
score; female is a binary gender variable; and athlete is a binary variable, which is one for
student athletes. What are your expectations for the coefficients in this equation? Which ones
are you unsure about?
58860_ch07_hr_220-261.indd 256
10/18/18 4:28 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
257
(ii)
Estimate the equation in part (i) and report the results in the usual form. What is the estimated
GPA differential between athletes and nonathletes? Is it statistically significant?
(iii) Drop sat from the model and reestimate the equation. Now, what is the estimated effect of being
an athlete? Discuss why the estimate is different than that obtained in part (ii).
(iv) In the model from part (i), allow the effect of being an athlete to differ by gender and test the
null hypothesis that there is no ceteris paribus difference between women athletes and women
nonathletes.
(v) Does the effect of sat on colgpa differ by gender? Justify your answer.
C5 In Problem 2 in Chapter 4, we added the return on the firm’s stock, ros, to a model explaining CEO
salary; ros turned out to be insignificant. Now, define a dummy variable, rosneg, which is equal to one
if ros , 0 and equal to zero if ros $ 0. Use CEOSAL1 to estimate the model
log 1 salary 2 5 b0 1 b1log 1 sales 2 1 b2roe 1 b3rosneg 1 u.
Discuss the interpretation and statistical significance of b^ 3.
C6 Use the data in SLEEP75 for this exercise. The equation of interest is
sleep 5 b0 1 b1totwrk 1 b2educ 1 b3age 1 b4age2 1 b5yngkid 1 u.
(i)
Estimate this equation separately for men and women and report the results in the usual form.
Are there notable differences in the two estimated equations?
(ii) Compute the Chow test for equality of the parameters in the sleep equation for men and women.
Use the form of the test that adds male and the interaction terms male∙totwrk, …, male∙yngkid
and uses the full set of observations. What are the relevant df for the test? Should you reject the
null at the 5% level?
(iii) Now, allow for a different intercept for males and females and determine whether the interaction
terms involving male are jointly significant.
(iv) Given the results from parts (ii) and (iii), what would be your final model?
C7 Use the data in WAGE1 for this exercise.
(i) Use equation (7.18) to estimate the gender differential when educ 5 12.5.
Compare this with the estimated differential when educ 5 0.
(ii) Run the regression used to obtain (7.18), but with female # 1 educ 2 12.5 2 replacing female∙educ.
How do you interpret the coefficient on female now?
(iii) Is the coefficient on female in part (ii) statistically significant? Compare this with (7.18) and
comment.
C8 Use the data in LOANAPP for this exercise. The binary variable to be explained is approve, which is
equal to one if a mortgage loan to an individual was approved. The key explanatory variable is white, a
dummy variable equal to one if the applicant was white. The other applicants in the data set are black
and Hispanic.
To test for discrimination in the mortgage loan market, a linear probability model can be used:
approve 5 b0 1 b1white 1 other factors.
(i)
If there is discrimination against minorities, and the appropriate factors have been controlled
for, what is the sign of b1?
(ii) Regress approve on white and report the results in the usual form. Interpret the coefficient on
white. Is it statistically significant? Is it practically large?
(iii) As controls, add the variables hrat, obrat, loanprc, unem, male, married, dep, sch, cosign,
chist, pubrec, mortlat1, mortlat2, and vr. What happens to the coefficient on white? Is there still
evidence of discrimination against nonwhites?
58860_ch07_hr_220-261.indd 257
10/18/18 4:28 PM
258
PART 1 Regression Analysis with Cross-Sectional Data
(iv)
(v)
Now, allow the effect of race to interact with the variable measuring other obligations as a
percentage of income (obrat). Is the interaction term significant?
Using the model from part (iv), what is the effect of being white on the probability of approval
when obrat 5 32, which is roughly the mean value in the sample? Obtain a 95% confidence
interval for this effect.
C9 There has been much interest in whether the presence of 401(k) pension plans, available to many U.S.
­workers, increases net savings. The data set 401KSUBS contains information on net financial assets (nettfa),
family income (inc), a binary variable for eligibility in a 401(k) plan (e401k), and several other variables.
(i) What fraction of the families in the sample are eligible for participation in a 401(k) plan?
(ii) Estimate a linear probability model explaining 401(k) eligibility in terms of income, age, and
gender. Include income and age in quadratic form, and report the results in the usual form.
(iii) Would you say that 401(k) eligibility is independent of income and age? What about gender?
Explain.
(iv) Obtain the fitted values from the linear probability model estimated in part (ii). Are any fitted
values negative or greater than one?
(v) Using the fitted values e401ki from part (iv), define e401ki 5 1 if e401ki $ .5 and e401ki 5 0 if
e401ki , .5. Out of 9,275 families, how many are predicted to be eligible for a 401(k) plan?
(vi) For the 5,638 families not eligible for a 401(k), what percentage of these are predicted not to
have a 401(k), using the predictor e401ki? For the 3,637 families eligible for a 401(k) plan,
what percentage are predicted to have one? (It is helpful if your econometrics package has a
“tabulate” command.)
(vii) The overall percent correctly predicted is about 64.9%. Do you think this is a complete
description of how well the model does, given your answers in part (vi)?
(viii) Add the variable pira as an explanatory variable to the linear probability model. Other things
equal, if a family has someone with an individual retirement account, how much higher is the
estimated probability that the family is eligible for a 401(k) plan? Is it statistically different
from zero at the 10% level?
C10 Use the data in NBASAL for this exercise.
(i) Estimate a linear regression model relating points per game to experience in the league and
position (guard, forward, or center). Include experience in quadratic form and use centers as the
base group. Report the results in the usual form.
(ii) Why do you not include all three position dummy variables in part (i)?
(iii) Holding experience fixed, does a guard score more than a center? How much more? Is the
difference statistically significant?
(iv) Now, add marital status to the equation. Holding position and experience fixed, are married
players more productive (based on points per game)?
(v) Add interactions of marital status with both experience variables. In this expanded model, is
there strong evidence that marital status affects points per game?
(vi) Estimate the model from part (iv) but use assists per game as the dependent variable. Are there
any notable differences from part (iv)? Discuss.
C11 Use the data in 401KSUBS for this exercise.
(i) Compute the average, standard deviation, minimum, and maximum values of nettfa in the
sample.
(ii) Test the hypothesis that average nettfa does not differ by 401(k) eligibility status; use a twosided alternative. What is the dollar amount of the estimated difference?
(iii) From part (ii) of Computer Exercise C9, it is clear that e401k is not exogenous in a simple
regression model; at a minimum, it changes by income and age. Estimate a multiple linear
regression model for nettfa that includes income, age, and e401k as explanatory variables. The
income and age variables should appear as quadratics. Now, what is the estimated dollar effect
of 401(k) eligibility?
58860_ch07_hr_220-261.indd 258
10/18/18 4:28 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
259
To the model estimated in part (iii), add the interactions e401k # 1 age 2 41 2 and
e401k # 1 age 2 41 2 2. Note that the average age in the sample is about 41, so that in the new
model, the coefficient on e401k is the estimated effect of 401(k) eligibility at the average age.
Which interaction term is significant?
(v) Comparing the estimates from parts (iii) and (iv), do the estimated effects of 401(k) eligibility at
age 41 differ much? Explain.
(vi) Now, drop the interaction terms from the model, but define five family size dummy variables:
fsize1, fsize2, fsize3, fsize4, and fsize5. The variable fsize5 is unity for families with five or more
members. Include the family size dummies in the model estimated from part (iii); be sure to
choose a base group. Are the family dummies significant at the 1% level?
(vii) Now, do a Chow test for the model
(iv)
nettfa 5 b0 1 b1inc 1 b2inc2 1 b3age 1 b4age2 1 b5e401k 1 u
across the five family size categories, allowing for intercept differences. The restricted sum of
squared residuals, SSRr, is obtained from part (vi) because that regression assumes all slopes are
the same. The unrestricted sum of squared residuals is SSRur 5 SSR1 1 SSR2 1 p 1 SSR5,
where SSRf is the sum of squared residuals for the equation estimated using only family
size f. You should convince yourself that there are 30 parameters in the unrestricted model
(5 intercepts plus 25 slopes) and 10 parameters in the restricted model (5 intercepts plus
5 slopes). Therefore, the number of restrictions being tested is q 5 20, and the df for the
unrestricted model is 9,275 2 30 5 9,245.
C12 Use the data set in BEAUTY, which contains a subset of the variables (but more usable observations
than in the regressions) reported by Hamermesh and Biddle (1994).
(i) Find the separate fractions of men and women that are classified as having above average looks.
Are more people rated as having above average or below average looks?
(ii) Test the null hypothesis that the population fractions of above-average-looking women and
men are the same. Report the one-sided p-value that the fraction is higher for women. (Hint:
Estimating a simple linear probability model is easiest.)
(iii) Now estimate the model
log 1 wage 2 5 b0 1 b1belavg 1 b2abvavg 1 u
separately for men and women, and report the results in the usual form. In both cases, interpret
the coefficient on belavg. Explain in words what the hypothesis H0: b1 5 0 against H1: b1 , 0
means, and find the p-values for men and women.
(iv) Is there convincing evidence that women with above average looks earn more than women with
average looks? Explain.
(v) For both men and women, add the explanatory variables educ, exper, exper2, union, goodhlth,
black, married, south, bigcity, smllcity, and service. Do the effects of the “looks” variables
change in important ways?
(vi) Use the SSR form of the Chow F statistic to test whether the slopes of the regression functions
in part (v) differ across men and women. Be sure to allow for an intercept shift under the null.
C13 Use the data in APPLE to answer this question.
(i) Define a binary variable as ecobuy 5 1 if ecolbs . 0 and ecobuy 5 0 if ecolbs 5 0. In other
words, ecobuy indicates whether, at the prices given, a family would buy any ecologically
friendly apples. What fraction of families claim they would buy ecolabeled apples?
(ii) Estimate the linear probability model
ecobuy 5 b0 1 b1ecoprc 1 b2 regprc 1 b3 faminc
1 b4 hhsize 1 b5 educ 1 b6 age 1 u,
and report the results in the usual form. Carefully interpret the coefficients on the price variables.
58860_ch07_hr_220-261.indd 259
10/18/18 4:28 PM
260
PART 1 Regression Analysis with Cross-Sectional Data
(iii) Are the nonprice variables jointly significant in the LPM? (Use the usual F statistic, even
though it is not valid when there is heteroskedasticity.) Which explanatory variable other than
the price variables seems to have the most important effect on the decision to buy ecolabeled
apples? Does this make sense to you?
(iv) In the model from part (ii), replace faminc with log(faminc). Which model fits the data better,
using faminc or log(faminc)? Interpret the coefficient on log(faminc).
(v) In the estimation in part (iv), how many estimated probabilities are negative? How many are
bigger than one? Should you be concerned?
(vi) For the estimation in part (iv), compute the percent correctly predicted for each outcome,
ecobuy 5 0 and ecobuy 5 1. Which outcome is best predicted by the model?
C14 Use the data in CHARITY to answer this question. The variable respond is a dummy variable equal to
one if a person responded with a contribution on the most recent mailing sent by a charitable organization. The variable resplast is a dummy variable equal to one if the person responded to the previous
mailing, avggift is the average of past gifts (in Dutch guilders), and propresp is the proportion of times
the person has responded to past mailings.
(i) Estimate a linear probability model relating respond to resplast and avggift. Report the results
in the usual form, and interpret the coefficient on resplast.
(ii) Does the average value of past gifts seem to affect the probability of responding?
(iii) Add the variable propresp to the model, and interpret its coefficient. (Be careful here: an
increase of one in propresp is the largest possible change.)
(iv) What happened to the coefficient on resplast when propresp was added to the regression? Does
this make sense?
(v) Add mailsyear, the number of mailings per year, to the model. How big is its estimated effect?
Why might this not be a good estimate of the causal effect of mailings on responding?
C15 Use the data in FERTIL2 to answer this question.
(i) Find the smallest and largest values of children in the sample. What is the average of children?
Does any woman have exactly the average number of children?
(ii) What percentage of women have electricity in the home?
(iii) Compute the average of children for those without electricity and do the same for those with
electricity. Comment on what you find. Test whether the population means are the same using a
simple regression.
(iv) From part (iii), can you infer that having electricity “causes” women to have fewer children?
Explain.
(v) Estimate a multiple regression model of the kind reported in equation (7.37), but add age2,
­urban, and the three religious affiliation dummies. How does the estimated effect of having
electricity compare with that in part (iii)? Is it still statistically significant?
(vi) To the equation in part (v), add an interaction between electric and educ. Is its coefficient
­statistically significant? What happens to the coefficient on electric?
(vii) The median and mode value for educ is 7. In the equation from part (vi), use the centered
­interaction term electric ? 1 educ 2 7 2 in place of electric ? educ. What happens to the coefficient on electric compared with part (vi)? Why? How does the coefficient on electric compare
with that in part (v)?
C16 Use the data in CATHOLIC to answer this question.
(i) In the entire sample, what percentage of the students attend a Catholic high school? What is the
average of math12 in the entire sample?
(ii) Run a simple regression of math12 on cathhs and report the results in the usual way. Interpret
what you have found.
58860_ch07_hr_220-261.indd 260
10/18/18 4:28 PM
CHAPTER 7 Multiple Regression Analysis with Qualitative Information
261
(iii) Now add the variables lfaminc, motheduc, and fatheduc to the regression from part (ii). How
many observations are used in the regression? What happens to the coefficient on cathhs, along
with its statistical significance?
(iv) Return to the simple regression of math12 on cathhs, but restrict the regression to observations
used in the multiple regression from part (iii). Do any important conclusions change?
(v) To the multiple regression in part (iii), add interactions between cathhs and each of the other
explanatory variables. Are the interaction terms individually or jointly significant?
(vi) What happens to the coefficient on cathhs in the regression from part (v). Explain why this
coefficient is not very interesting.
(vii) Compute the average partial effect of cathhs in the model estimated in part (v). How does it
compare with the coefficients on cathhs in parts (iii) and (v)?
C17 Use the data in JTRAIN98 to answer this question. The variable unem98 is a binary variable indicating
whether a worker was unemployed in 1998. It can be used to measure the effectiveness of the job training program in reducing the probability of being unemployed.
(i) What percentage of workers was unemployed in 1998, after the job training program? How
does this compare with the unemployment rate in 1996?
(ii) Run the simple regression unem98 on train. How do you interepret the coefficient on train? Is it
statistically significant? Does it make sense to you?
(iii) Add to the regression in part (ii) the explanatory variables earn96, educ, age, a n d married.
Now interpret the estimated training effect. Why does it differ so much from that in part (ii)?
(iv) Now perform full regression adjustment by running a regression with a full set of interactions,
where all variables (except the training indicator) are centered around their sample means:
unem98i on traini, earn96i, educi, agei, married, traini ? 1 earn96i 2 earn96),
traini ? 1 educi 2 educ), traini ? 1 agei 2 age), traini ? 1 marriedi 2 married).
This regression uses all of the data. What ha ppens to the estimated average treatment effect of train
compared with part (iii). Does its standard error change much?
(v) Are the interaction terms in part (iv) jointly significant?
(vi) Verify that you obtain exactly the same average treatment effect if you run two separate
regressions and use the formula in equation (7.43). That is, run two separate regressions for the
control and treated groups, obtain the fitted values unem98i102 and unem98i112 for everyone in the
sample, and then compute
t^ ura 5 n 21 a 3 unem98i112 2 unem98i102 4 .
n
i51
Check this with the coefficient on train in part (iv). Which approach is more convenient for obtaining
a standard error?
58860_ch07_hr_220-261.indd 261
10/18/18 4:28 PM
chapter
8
Heteroskedasticity
T
he homoskedasticity assumption, introduced in Chapter 3 for multiple regression, states that
the variance of the unobserved error, u, conditional on the explanatory variables, is constant.
­Homoskedasticity fails whenever the variance of the unobserved factors changes across different segments of the population, where the segments are determined by the different values of the
explanatory variables. For example, in a savings equation, heteroskedasticity is present if the variance
of the unobserved factors affecting savings increases with income.
In Chapters 4 and 5, we saw that homoskedasticity is needed to justify the usual t tests, F tests,
and confidence intervals for OLS estimation of the linear regression model, even with large sample
sizes. In this chapter, we discuss the available remedies when heteroskedasticity occurs, and we also
show how to test for its presence. We begin by briefly reviewing the consequences of heteroskedasticity for ordinary least squares estimation.
8-1 Consequences of Heteroskedasticity for OLS
Consider again the multiple linear regression model:
y 5 b0 1 b1x1 1 b2x2 1 p 1 bkxk 1 u.
262
[8.1]
In Chapter 3, we proved unbiasedness of the OLS estimators b^ 0, b^ 1, b^ 2, . . . , b^ k under the first four
Gauss-Markov assumptions, MLR.1 through MLR.4. In Chapter 5, we showed that the same four
assumptions imply consistency of OLS. The homoskedasticity assumption MLR.5, stated in terms of the
error variance as Var 1 u 0 x1, x2, . . . , xk 2 5 s2, played no role in showing whether OLS was unbiased or
58860_ch08_hr_262-293.indd 262
10/18/18 4:28 PM
CHAPTER 8 Heteroskedasticity
263
consistent. It is important to remember that heteroskedasticity does not cause bias or inconsistency in the
OLS estimators of the bj, whereas something like omitting an important variable would have this effect.
The interpretation of our goodness-of-fit measures, R2 and R2, is also unaffected by the presence of heteroskedasticity. Why? Recall from Section 6-3 that the usual R-squared and the adjusted
R-squared are different ways of estimating the population R-squared, which is simply 1 2 s2u/s2y ,
where s2u is the population error variance and s2y is the population variance of y. The key point is
that because both variances in the population R-squared are unconditional variances, the population
R-squared is unaffected by the presence of heteroskedasticity in Var 1 u 0 x1, . . . , xk 2 . Further, SSR/n
consistently estimates s2u, and SST/n consistently estimates s2y , whether or not Var 1 u 0 x1, . . . , xk 2
is constant. The same is true when we use the degrees of freedom adjustments. Therefore, R2 and
R2 are both consistent estimators of the population R-squared whether or not the homoskedasticity
­assumption holds.
If heteroskedasticity does not cause bias or inconsistency in the OLS estimators, why did we
introduce it as one of the Gauss-Markov assumptions? Recall from Chapter 3 that the estimators
of the variances, Var 1 b^ j 2 , are biased without the homoskedasticity assumption. Because the OLS
­standard errors are based directly on these variances, they are no longer valid for constructing confidence intervals and t statistics. The usual OLS t statistics do not have t distributions in the presence
of heteroskedasticity, and the problem is not resolved by using large sample sizes. We will see this
explicitly for the simple regression case in the next section, where we derive the variance of the
OLS slope estimator under heteroskedasticity and propose a valid estimator in the presence of heteroskedasticity. Similarly, F statistics are no longer F distributed, and the LM statistic no longer has
an asymptotic chi-square distribution. In summary, the statistics we used to test hypotheses under the
Gauss-Markov assumptions are not valid in the presence of heteroskedasticity.
We also know that the Gauss-Markov Theorem, which says that OLS is best linear unbiased,
relies crucially on the homoskedasticity assumption. If Var 1 u 0 x 2 is not constant, OLS is no longer
BLUE. In addition, OLS is no longer asymptotically efficient in the class of estimators described in
Theorem 5.3. As we will see in Section 8-4, it is possible to find estimators that are more efficient
than OLS in the presence of heteroskedasticity (although it requires knowing the form of the heteroskedasticity). With relatively large sample sizes, it might not be so important to obtain an efficient
estimator. In the next section, we show how the usual OLS test statistics can be modified so that they
are valid, at least asymptotically.
8-2 Heteroskedasticity-Robust Inference after OLS Estimation
Because testing hypotheses is such an important component of any econometric analysis and the usual
OLS inference is generally faulty in the presence of heteroskedasticity, we must decide if we should
entirely abandon OLS. Fortunately, OLS is still useful. In the last two decades, econometricians have
learned how to adjust standard errors and t, F, and LM statistics so that they are valid in the presence
of heteroskedasticity of unknown form. This is very convenient because it means we can report new
statistics that work regardless of the kind of heteroskedasticity present in the population. The methods in
this section are known as heteroskedasticity-robust procedures because they are valid—at least in large
samples—whether or not the errors have constant variance, and we do not need to know which is the case.
We begin by sketching how the variances, Var 1 b^ j 2 , can be estimated in the presence of heteroskedasticity. A careful derivation of the theory is well beyond the scope of this text, but the application of
heteroskedasticity-robust methods is very easy now because many statistics and econometrics packages compute these statistics as an option.
First, consider the model with a single independent variable, where we include an i subscript for
emphasis:
yi 5 b0 1 b1xi 1 ui.
58860_ch08_hr_262-293.indd 263
10/18/18 4:28 PM
264
PART 1 Regression Analysis with Cross-Sectional Data
We assume throughout that the first four Gauss-Markov assumptions hold. If the errors contain
­heteroskedasticity, then
Var 1 ui 0 xi 2 5 s2i ,
where we put an i subscript on s2 to indicate that the variance of the error depends upon the particular
value of xi.
Write the OLS estimator as
a 1 xi 2 x 2 ui
n
b^ 1 5 b1 1
i51
n
2
a 1 xi 2 x 2
.
i51
Under Assumptions MLR.1 through MLR.4 (that is, without the homoskedasticity assumption), and
conditioning on the values xi in the sample, we can use the same arguments from Chapter 2 to show
that
2 2
a 1 xi 2 x 2 si
n
Var 1 b^ 1 2 5
i51
SST2x
,
[8.2]
n
where SSTx 5 g i51 1 xi 2 x 2 2 is the total sum of squares of the xi. When s2i 5 s2 for all i, this formula
reduces to the usual form, s2/SSTx. Equation (8.2) explicitly shows that, for the simple regression
case, the variance formula derived under homoskedasticity is no longer valid when heteroskedasticity
is present.
Because the standard error of b^ 1 is based directly on estimating Var 1 b^ 1 2 , we need a way to
­estimate equation (8.2) when heteroskedasticity is present. White (1980) showed how this can be
done. Let u^ i denote the OLS residuals from the initial regression of y on x. Then, a valid estimator of
Var 1 b^ 1 2 , for heteroskedasticity of any form (including homoskedasticity), is
2 2
a 1 xi 2 x 2 u^ i
n
i51
SST2x
[8.3]
,
which is easily computed from the data after the OLS regression.
In what sense is (8.3) a valid estimator of Var 1 b^ 1 2 ? This is pretty subtle. Briefly, it can be
shown that when equation (8.3) is multiplied by the sample size n, it converges in probability to
E 3 1 xi 2 mx 2 2u2i 4 / 1 s2x 2 2, which is the probability limit of n times (8.2). Ultimately, this is what is necessary for justifying the use of standard errors to construct confidence intervals and t statistics. The law
of large numbers and the central limit theorem play key roles in establishing these convergences. You
can refer to White’s original paper for details, but that paper is quite technical. See also Wooldridge
(2010, Chapter 4).
A similar formula works in the general multiple regression model
y 5 b0 1 b1x1 1 p 1 bkxk 1 u.
It can be shown that a valid estimator of Var 1 b^ j 2 , under Assumptions MLR.1 through MLR.4, is
a r^ iju^ i
n
Var 1 b^ j 2 5
58860_ch08_hr_262-293.indd 264
2 2
i51
SSR2j
,
[8.4]
10/18/18 4:28 PM
CHAPTER 8 Heteroskedasticity
265
where r^ ij denotes the ith residual from regressing xj on all other independent variables, and SSRj is the
sum of squared residuals from this regression (see Section 3-2 for the partialling out representation of the
OLS estimates). The square root of the quantity in (8.4) is called the heteroskedasticity-robust ­standard
error for b^ j. In econometrics, these robust standard errors are usually attributed to White (1980). Earlier
works in statistics, notably those by Eicker (1967) and Huber (1967), pointed to the possibility of obtaining such robust standard errors. In applied work, these are sometimes called White, Huber, or Eicker
standard errors (or some hyphenated combination of these names). We will just refer to them as heteroskedasticity-robust standard errors, or even just robust standard errors when the context is clear.
Sometimes, as a degrees of freedom correction, (8.4) is multiplied by n/ 1 n 2 k 2 1 2 before taking
the square root. The reasoning for this adjustment is that, if the squared OLS residuals u^ 2i were the same for
all observations i—the strongest possible form of homoskedasticity in a sample—we would get the usual
OLS standard errors. Other modifications of (8.4) are studied in MacKinnon and White (1985). Because
all forms have only asymptotic justification and they are asymptotically equivalent, no form is uniformly
preferred above all others. Typically, we use whatever form is computed by the regression package at hand.
Once heteroskedasticity-robust standard errors are obtained, it is simple to construct a
heteroskedasticity-robust t statistic. Recall that the general form of the t statistic is
estimate 2 hypothesized value
.
[8.5]
standard error
Because we are still using the OLS estimates and we have chosen the hypothesized value ahead of
time, the only difference between the usual OLS t statistic and the heteroskedasticity-robust t statistic
is in how the standard error in the denominator is computed.
The term SSRj in equation (8.4) can be replaced with SSTj 1 1 2 R2j 2 , where SSTj is the total
sum of squares of xj and R2j is the usual R-squared from regressing xj on all other explanatory variables. [We implicitly used this equivalence in deriving equation (3.51).] Consequently, little sample
variation in xj, or a strong linear relationship between xj and the other explanatory variables—that is,
­multicollinearity—can cause the heteroskedasticity-robust standard errors to be large. We discussed
these issues with the usual OLS standard errors in Section 3-4.
t5
Example 8.1
Log Wage Equation with Heteroskedasticity-Robust
Standard Errors
We estimate the model in Example 7.6, but we report the heteroskedasticity-robust standard errors
along with the usual OLS standard errors. Some of the estimates are reported to more digits so that we
can compare the usual standard errors with the heteroskedasticity-robust standard errors:
log 1 wage 2 5 .321 1 .213 marrmale 2 .198 marrfem 2 .110 singfem
1 .100 2 1 .055 2
1 .058 2
1 .056 2
3 .109 4 3 .057 4
3 .058 4
3 .057 4
1 .0789 educ 1 .0268 exper 2 .00054 exper2
1 .0067 2
1 .0052 2
1 .00011 2
3 .0074 4
3 .0051 4
3 .00011 4
[8.6]
1 .0291 tenure 2 .00053 tenure2
1 .0068 2
1 .00023 2
3 .0069 4
3 .00024 4
n 5 526, R2 5 .461.
The usual OLS standard errors are in parentheses, ( ), below the corresponding OLS estimate, and
the heteroskedasticity-robust standard errors are in brackets, [ ]. The numbers in brackets are the only
new things, as the equation is still estimated by OLS.
58860_ch08_hr_262-293.indd 265
10/18/18 4:28 PM
266
PART 1 Regression Analysis with Cross-Sectional Data
Several things are apparent from equation (8.6). First, in this particular application, any variable
that was statistically significant using the usual t statistic is still statistically significant using the
heteroskedasticity-robust t statistic. This occurs because the two sets of standard errors are not very
different. (The associated p-values will differ slightly because the robust t statistics are not identical
to the usual, nonrobust t statistics.) The largest relative change in standard errors is for the coefficient
on educ: the usual standard error is .0067, and the robust standard error is .0074. Still, the robust standard error implies a robust t statistic above 10.
Equation (8.6) also shows that the robust standard errors can be either larger or smaller than the
usual standard errors. For example, the robust standard error on exper is .0051, whereas the usual
standard error is .0055. We do not know which will be larger ahead of time. As an empirical matter,
the robust standard errors are often found to be larger than the usual standard errors.
Before leaving this example, we must emphasize that we do not know, at this point, whether
heteroskedasticity is even present in the population model underlying equation (8.6). All we have
done is report, along with the usual standard errors, those that are valid (asymptotically) whether or
not heteroskedasticity is present. We can see that no important conclusions are overturned by using
the robust standard errors in this example. This often happens in applied work, but in other cases, the
differences between the usual and robust standard errors are much larger. As an example of where the
differences are substantial, see Computer Exercise C2.
At this point, you may be asking the following question: if the heteroskedasticity-robust standard
errors are valid more often than the usual OLS standard errors, why do we bother with the usual standard errors at all? This is a sensible question. One reason the usual standard errors are still used in crosssectional work is that, if the homoskedasticity assumption holds and the errors are normally distributed,
then the usual t statistics have exact t distributions, regardless of the sample size (see Chapter 4).
The robust standard errors and robust t statistics are justified only as the sample size becomes large,
even if the CLM assumptions are true. With small sample sizes, the robust t statistics can have distributions that are not very close to the t distribution, and that could throw off our inference.
In large sample sizes, we can make a case for always reporting only the heteroskedasticity-robust
standard errors in cross-sectional applications, and this practice is being followed more and more in
applied work. It is also common to report both standard errors, as in equation (8.6), so that a reader
can determine whether any conclusions are sensitive to the standard error in use.
It is also possible to obtain F and LM statistics that are robust to heteroskedasticity of an unknown,
arbitrary form. The heteroskedasticity-robust F statistic (or a simple transformation of it) is also called
a heteroskedasticity-robust Wald statistic. A general treatment of the Wald statistic requires matrix algebra and is sketched in Advanced Treatment E; see Wooldridge (2010, Chapter 4) for a more detailed
treatment. Nevertheless, using heteroskedasticity-robust statistics for multiple exclusion restrictions is
straightforward because many econometrics packages now compute such statistics routinely.
Example 8.2
Heteroskedasticity-Robust F Statistic
Using the data for the spring semester in GPA3, we estimate the following equation:
cumgpa 5 1.47 1 .00114 sat 2 .00857 hsperc 1 .00250 tothrs
1 .23 2 1 .00018 2
1 .00124 2
1 .00073 2
3 .22 4 3 .00019 4
3 .00140 4
3 .00073 4
1 .303 female 2 .128 black 2 .059 white
1 .059 2
1 .147 2
1 .141 2
3 .059 4
3 .118 4
3 .110 4
[8.7]
n 5 366, R2 5 .4006, R2 5 .3905.
58860_ch08_hr_262-293.indd 266
10/18/18 4:28 PM
CHAPTER 8 Heteroskedasticity
267
Again, the differences between the usual standard errors and the heteroskedasticity-robust standard
errors are not very big, and use of the robust t statistics does not change the statistical significance
of any independent variable. Joint significance tests are not much affected either. Suppose we wish
to test the null hypothesis that, after the other factors are controlled for, there are no differences in
cumgpa by race. This is stated as H0: bblack 5 0, bwhite 5 0. The usual F statistic is easily obtained,
once we have the R-squared from the restricted model; this turns out to be .3983. The F statistic is
then 3 1 .4006 2 .3983 2 / 1 1 2 .4006 2 4 1 359/2 2 < .69. If heteroskedasticity is present, this version of
the test is invalid. The heteroskedasticity-robust version has no simple form, but it can be computed
using certain statistical packages. The value of the heteroskedasticity-robust F statistic turns out to
be .75, which differs only slightly from the nonrobust version. The p-value for the robust test is .474,
which is not close to standard significance levels. We fail to reject the null hypothesis using either test.
Because the usual sum of squared residuals form of the F statistic is not valid under heteroskedasticity, we must be careful in computing a Chow test of common coefficients across two groups. The
form of the statistic in equation (7.24) is not valid if heteroskedasticity is present, including the simple
case where the error variance differs across the two groups. Instead, we can obtain a heteroskedasticityrobust Chow test by including a dummy variable distinguishing the two groups along with interactions
between that dummy variable and all other explanatory variables. We can then test whether there is no
difference in the two regression functions—by testing that the coefficients on the dummy variable and
all interactions are zero—or just test whether the slopes are all the same, in which case we leave the
coefficient on the dummy variable unrestricted. See Computer Exercise C14 for an example.
8-2a Computing Heteroskedasticity-Robust LM Tests
Not all regression packages compute F statistics
that are robust to heteroskedasticity. Therefore, it is
sometimes convenient to have a way of obtaining a
Evaluate the following statement: The
test of multiple exclusion restrictions that is robust
heteroskedasticity-robust standard errors
are always bigger than the usual standard
to heteroskedasticity and does not require a particuerrors.
lar kind of econometric software. It turns out that a
heteroskedasticity-robust LM statistic is easily
­obtained using virtually any regression package.
To illustrate computation of the robust LM statistic, consider the model
G o i n g F u rt h e r 8 . 1
y 5 b0 1 b1x1 1 b2x2 1 b3x3 1 b4x4 1 b5x5 1 u,
and suppose we would like to test H0: b4 5 0, b5 5 0. To obtain the usual LM statistic, we would
first estimate the restricted model (that is, the model without x4 and x5) to obtain the residuals, |
u.
Then, we would regress |
u on all of the independent variables and the LM 5 n # R2u& , where R2u& is the
usual R-squared from this regression.
Obtaining a version that is robust to heteroskedasticity requires more work. One way to compute
the statistic requires only OLS regressions. We need the residuals, say, |
r1 , from the regression of
x4 on x1, x2, x3. Also, we need the residuals, say, |
r2, from the regression of x5 on x1, x2, x3. Thus, we
regress each of the independent variables excluded under the null on all of the included independent
variables. We keep the residuals each time. The final step appears odd, but it is, after all, just a computational device. Run the regression of
1 on |
r 1|
u, |
r 2|
u,
[8.8]
without an intercept. Yes, we actually define a dependent variable equal to the value one for all
observations. We regress this onto the products |
r1 |
u and |
r 2|
u . The robust LM statistic turns out to be
n 2 SSR1, where SSR1 is just the usual sum of squared residuals from regression (8.8).
58860_ch08_hr_262-293.indd 267
10/18/18 4:28 PM
268
PART 1 Regression Analysis with Cross-Sectional Data
The reason this works is somewhat technical. Basically, this is doing for the LM test what the
robust standard errors do for the t test. [See Wooldridge (1991b) or Davidson and MacKinnon (1993)
for a more detailed discussion.]
We now summarize the computation of the heteroskedasticity-robust LM statistic in the general case.
A Heteroskedasticity-Robust LM Statistic:
1. Obtain the residuals |
u from the restricted model.
2. Regress each of the independent variables excluded under the null on all of the included independent variables; if there are q excluded variables, this leads to q sets of residuals 1 |
r 1, |
r2 , . . . , |
rq 2 .
|
|
3. Find the products of each rj and u (for all observations).
4. Run the regression of 1 on |
r1|
u, |
r 2|
u, . . . , |
r q|
u , without an intercept. The heteroskedasticity-robust
LM statistic is n 2 SSR1, where SSR1 is just the usual sum of squared residuals from this final
regression. Under H0, LM is distributed approximately as x2q.
Once the robust LM statistic is obtained, the rejection rule and computation of p-values are the same
as for the usual LM statistic in Section 5-2.
Example 8.3
Heteroskedasticity-Robust LM Statistic
We use the data in CRIME1 to test whether the average sentence length served for past convictions
affects the number of arrests in the current year (1986). The estimated model is
narr86 5 .561 2 .136 pcnv 1 .0178 avgsen 2 .00052 avgsen2
1 .036 2 1 .040 2
1 .0097 2
1 .00030 2
3 .040 4 3 .034 4
3 .0101 4
3 .00021 4
2 .0394 ptime86 2 .0505 qemp86 2 .00148 inc86
1 .0087 2
1 .0144 2
1 .00034 2
3 .0062 4
3 .0142 4
3 .00023 4
[8.9]
1 .325 black 1 .193 hispan
1 .045 2
1 .040 2
3 .058 4
3 .040 4
n 5 2,725, R2 5 .0728.
In this example, there are more substantial differences between some of the usual standard errors
and the robust standard errors. For example, the usual t statistic on avgsen2 is about −1.73, while the
robust t statistic is about −2.48. Thus, avgsen2 is more significant using the robust standard error.
The effect of avgsen on narr86 is somewhat difficult to reconcile. Because the relationship
is quadratic, we can figure out where avgsen has a positive effect on narr86 and where the effect
becomes negative. The turning point is .0178/[2(.00052)] L 17.12; recall that this is measured in
months. Literally, this means that narr86 is positively related to avgsen when avgsen is less than 17
months; then avgsen has the expected deterrent effect after 17 months.
To see whether average sentence length has a statistically significant effect on narr86, we must
test the joint hypothesis H0: bavgsen 5 0, bavgsen2 5 0. Using the usual LM statistic (see Section 5-2), we
obtain LM 5 3.54; in a chi-square distribution with two df, this yields a p-value 5 .170. Thus, we do
not reject H0 at even the 15% level. The heteroskedasticity-robust LM statistic is LM 5 4.00 (rounded
to two decimal places), with a p-value 5 .135. This is still not very strong evidence against H0; avgsen
does not appear to have a strong effect on narr86. [Incidentally, when avgsen appears alone in (8.9),
that is, without the quadratic term, its usual t statistic is .658, and its robust t statistic is .592.]
58860_ch08_hr_262-293.indd 268
10/18/18 4:28 PM
CHAPTER 8 Heteroskedasticity
269
8-3 Testing for Heteroskedasticity
The heteroskedasticity-robust standard errors provide a simple method for computing t statistics that
are asymptotically t distributed whether or not heteroskedasticity is present. We have also seen that
heteroskedasticity-robust F and LM statistics are available. Implementing these tests does not require
knowing whether or not heteroskedasticity is present. Nevertheless, there are still some good reasons
for having simple tests that can detect its presence. First, as we mentioned in the previous section,
the usual t statistics have exact t distributions under the classical linear model assumptions. For this
reason, many economists still prefer to see the usual OLS standard errors and test statistics reported,
unless there is evidence of heteroskedasticity. Second, if heteroskedasticity is present, the OLS estimator is no longer the best linear unbiased estimator. As we will see in Section 8-4, it is possible to
obtain a better estimator than OLS when the form of heteroskedasticity is known.
Many tests for heteroskedasticity have been suggested over the years. Some of them, while having the ability to detect heteroskedasticity, do not directly test the assumption that the variance of
the error does not depend upon the independent variables. We will restrict ourselves to more modern
tests, which detect the kind of heteroskedasticity that invalidates the usual OLS statistics. This also
has the benefit of putting all tests in the same framework.
As usual, we start with the linear model
y 5 b0 1 b1x1 1 b2x2 1 p 1 bkxk 1 u,
[8.10]
where Assumptions MLR.1 through MLR.4 are maintained in this section. In particular, we assume
that E 1 u 0 x1, x2, . . . , xk 2 5 0, so that OLS is unbiased and consistent.
We take the null hypothesis to be that Assumption MLR.5 is true:
H0: Var 1 u 0 x1, x2, . . . , xk 2 5 s2.
[8.11]
That is, we assume that the ideal assumption of homoskedasticity holds, and we require the data
to tell us otherwise. If we cannot reject (8.11) at a sufficiently small significance level, we usually
conclude that heteroskedasticity is not a problem. However, remember that we never accept H0; we
simply fail to reject it.
Because we are assuming that u has a zero conditional expectation, Var 1 u 0 x 2 5 E 1 u2 0 x 2 , and so
the null hypothesis of homoskedasticity is equivalent to
H0: E 1 u2 0 x1, x2, . . . , xk 2 5 E 1 u2 2 5 s2.
This shows that, in order to test for violation of the homoskedasticity assumption, we want to test
whether u2 is related (in expected value) to one or more of the explanatory variables. If H0 is false, the
expected value of u2, given the independent variables, can be virtually any function of the xj. A simple
approach is to assume a linear function:
u2 5 d0 1 d1x1 1 d2x2 1 p 1 dkxk 1 v,
[8.12]
where v is an error term with mean zero given the xj. Pay close attention to the dependent variable in
this equation: it is the square of the error in the original regression equation, (8.10). The null hypothesis of homoskedasticity is
H0: d1 5 d2 5 p 5 dk 5 0.
[8.13]
Under the null hypothesis, it is often reasonable to assume that the error in (8.12), v, is independent
of x1, x2, . . . , xk. Then, we know from Section 5-2 that either the F or LM statistics for the overall significance of the independent variables in explaining u2 can be used to test (8.13). Both statistics would
have asymptotic justification, even though u2 cannot be normally distributed. (For example, if u is
normally distributed, then u2/s2 is distributed as x21.) If we could observe the u2 in the sample, then
we could easily compute this statistic by running the OLS regression of u2 on x1, x2, . . . , xk, using all
n observations.
58860_ch08_hr_262-293.indd 269
10/18/18 4:28 PM
270
PART 1 Regression Analysis with Cross-Sectional Data
As we have emphasized before, we never know the actual errors in the population model, but we
do have estimates of them: the OLS residual, u^ i, is an estimate of the error ui for observation i. Thus,
we can estimate the equation
u^ 2 5 d0 1 d1x1 1 d2x2 1 p 1 dkxk 1 error
[8.14]
and compute the F or LM statistics for the joint significance of x1, . . . , xk. It turns out that using the
OLS residuals in place of the errors does not affect the large sample distribution of the F or LM statistics, although showing this is pretty complicated.
The F and LM statistics both depend on the R-squared from regression (8.14); call this R2u^ 2 to distinguish it from the R-squared in estimating equation (8.10). Then, the F statistic is
F5
R2u^ 2/k
,
1 1 2 R2u^ 2 2 / 1 n 2 k 2 1 2
[8.15]
where k is the number of regressors in (8.14); this is the same number of independent variables in
(8.10). Computing (8.15) by hand is rarely necessary, because most regression packages automatically compute the F statistic for overall significance of a regression. This F statistic has (approximately) an Fk, n2k21 distribution under the null hypothesis of homoskedasticity.
The LM statistic for heteroskedasticity is just the sample size times the R-squared from (8.14):
LM 5 n # R2u^ 2.
[8.16]
Under the null hypothesis, LM is distributed asymptotically as x2k . This is also very easy to obtain
after running regression (8.14).
The LM version of the test is typically called the Breusch-Pagan test for heteroskedasticity
(BP test). Breusch and Pagan (1979) suggested a different form of the test that assumes the errors are
normally distributed. Koenker (1981) suggested the form of the LM statistic in (8.16), and it is generally preferred due to its greater applicability.
We summarize the steps for testing for heteroskedasticity using the BP test:
The Breusch-Pagan Test for Heteroskedasticity:
1. Estimate the model (8.10) by OLS, as usual. Obtain the squared OLS residuals, u^ 2 (one for each
observation).
2. Run the regression in (8.14). Keep the R-squared from this regression, R2u^ 2 .
3. Form either the F statistic or the LM statistic and compute the p-value (using the Fk, n2k21 distribution in the former case and the x2k distribution in the latter case). If the p-value is sufficiently small, that is, below the chosen significance level, then we reject the null hypothesis of
homoskedasticity.
If the BP test results in a small enough p-value, some corrective measure should be taken. One
possibility is to just use the heteroskedasticity-robust standard errors and test statistics discussed in
the previous section. Another possibility is discussed in Section 8-4.
Example 8.4
Heteroskedasticity in Housing Price Equations
We use the data in HPRICE1 to test for heteroskedasticity in a simple housing price equation. The
estimated equation using the levels of all variables is
price 5 221.77 1 .00207 lotsize 1 .123 sqrft 1 13.85 bdrms
1 29.48 2 1 .00064 2
1 .013 2
1 9.01 2
[8.17]
2
n 5 88, R 5 .672.
58860_ch08_hr_262-293.indd 270
10/18/18 4:28 PM
CHAPTER 8 Heteroskedasticity
271
This equation tells us nothing about whether the error in the population model is heteroskedastic.
We need to regress the squared OLS residuals on the independent variables. The R-squared from the
regression of u^ 2 on lotsize, sqrft, and bdrms is R2u^ 2 5 .1601. With n = 88 and k 5 3, this produces an
F statistic for significance of the independent variables of F 5 3 .1601/ 1 1 2 .1601 2 4 1 84/3 2 < 5.34.
The associated p-value is .002, which is strong evidence against the null. The LM statistic is
88 1 .1601 2 < 14.09; this gives a p-value < .0028 (using the x23 distribution), giving essentially the
same conclusion as the F statistic. This means that the usual standard errors reported in (8.17) are not
reliable.
In Chapter 6, we mentioned that one benefit of using the logarithmic functional form for the
dependent variable is that heteroskedasticity is often reduced. In the current application, let us put
price, lotsize, and sqrft in logarithmic form, so that the elasticities of price, with respect to lotsize and
sqrft, are constant. The estimated equation is
log 1 price 2 5 21.30 1 .168 log 1 lotsize 2 1 .700 log 1 sqrft 2 1 0.37 bdrms
1 .65 2 1 .038 2
1 .093 2
1 .028 2
[8.18]
2
n 5 88, R 5 .643.
Regressing the squared OLS residuals from this regression on log(lotsize), log(sqrft), and bdrms gives
R2u^ 2 5 .0480. Thus, F 5 1.41 1 p-value 5 .245 2 , and LM 5 4.22 1 p-value 5 .239 2 . Therefore, we
fail to reject the null hypothesis of homoskedasticity in the model with the logarithmic functional
forms. The occurrence of less heteroskedasticity with the dependent variable in logarithmic form has
been noticed in many empirical applications.
G o i n g F u rt h e r 8 . 2
Consider wage equation (7.11), where
you think that the conditional variance of
log(wage) does not depend on educ, exper,
or tenure. However, you are worried that
the variance of log(wage) differs across the
four demographic groups of married males,
married females, single males, and single
females. What regression would you run to
test for heteroskedasticity? What are the
degrees of freedom in the F test?
If we suspect that heteroskedasticity depends
only upon certain independent variables, we can
easily modify the Breusch-Pagan test: we simply
regress u^ 2 on whatever independent variables we
choose and carry out the appropriate F or LM test.
Remember that the appropriate degrees of freedom
depends upon the number of independent variables
in the regression with u^ 2 as the dependent variable;
the number of independent variables showing up in
equation (8.10) is irrelevant.
If the squared residuals are regressed on only a
single independent variable, the test for heteroskedasticity is just the usual t statistic on the variable. A
significant t statistic suggests that heteroskedasticity
is a problem.
8-3a The White Test for Heteroskedasticity
In Chapter 5, we showed that the usual OLS standard errors and test statistics are asymptotically
valid, provided all of the Gauss-Markov assumptions hold. It turns out that the homoskedasticity
assumption, Var 1 u1 0 x1, . . . , xk 2 5 s2, can be replaced with the weaker assumption that the squared
error, u2, is uncorrelated with all the independent variables 1 xj 2 , the squares of the independent variables 1 x2j 2 , and all the cross products 1 xj xh for j 2 h 2 . This observation motivated White (1980) to
propose a test for heteroskedasticity that adds the squares and cross products of all the independent
variables to equation (8.14). The test is explicitly intended to test for forms of heteroskedasticity that
invalidate the usual OLS standard errors and test statistics.
58860_ch08_hr_262-293.indd 271
10/18/18 4:29 PM
272
PART 1 Regression Analysis with Cross-Sectional Data
When the model contains k 5 3 independent variables, the White test is based on an
estimation of
u^ 2 5 d0 1 d1x1 1 d2x2 1 d3x3 1 d4x21 1 d5x22 1 d6x23
1 d7x1x2 1 d8x1x3 1 d9x2x3 1 error.
[8.19]
Compared with the Breusch-Pagan test, this equation has six more regressors. The White test for
heteroskedasticity is the LM statistic for testing that all of the dj in equation (8.19) are zero, except
for the intercept. Thus, nine restrictions are being tested in this case. We can also use an F test of this
hypothesis; both tests have asymptotic justification.
With only three independent variables in the original model, equation (8.19) has nine independent variables. With six independent variables in the original model, the White regression would generally involve 27 regressors (unless some are redundant). This abundance of regressors is a weakness
in the pure form of the White test: it uses many degrees of freedom for models with just a moderate
number of independent variables.
It is possible to obtain a test that is easier to implement than the White test and more conserving
on degrees of freedom. To create the test, recall that the difference between the White and BreuschPagan tests is that the former includes the squares and cross products of the independent variables.
We can preserve the spirit of the White test while conserving on degrees of freedom by using the OLS
fitted values in a test for heteroskedasticity. Remember that the fitted values are defined, for each
observation i, by
y^ i 5 b^ 0 1 b^ 1xi1 1 b^ 2xi2 1 p 1 b^ kxik.
These are just linear functions of the independent variables. If we square the fitted values, we get a
particular function of all the squares and cross products of the independent variables. This suggests
testing for heteroskedasticity by estimating the equation
u^ 2 5 d0 1 d1y^ 1 d2y^ 2 1 error,
[8.20]
where y^ stands for the fitted values. It is important not to confuse y^ and y in this equation. We use the
fitted values because they are functions of the independent variables (and the estimated parameters);
using y in (8.20) does not produce a valid test for heteroskedasticity.
We can use the F or LM statistic for the null hypothesis H0: d1 5 0, d2 5 0 in equation (8.20).
This results in two restrictions in testing the null of homoskedasticity, regardless of the number of
independent variables in the original model. Conserving on degrees of freedom in this way is often a
good idea, and it also makes the test easy to implement.
Because y^ is an estimate of the expected value of y, given the xj, using (8.20) to test for heteroskedasticity is useful in cases where the variance is thought to change with the level of the expected
value, E 1 y 0 x 2 . The test from (8.20) can be viewed as a special case of the White test, as equation
(8.20) can be shown to impose restrictions on the parameters in equation (8.19).
A Special Case of the White Test for Heteroskedasticity:
1. Estimate the model (8.10) by OLS, as usual. Obtain the OLS residuals u^ and the fitted values y^ .
Compute the squared OLS residuals u^ 2 and the squared fitted values y^ 2.
2. Run the regression in equation (8.20). Keep the R-squared from this regression, R2u^2.
3. Form either the F or LM statistic and compute the p-value (using the F2, n23 distribution in the
former case and the x22 distribution in the latter case).
58860_ch08_hr_262-293.indd 272
10/18/18 4:29 PM
CHAPTER 8 Heteroskedasticity
Example 8.5
273
pecial Form of the White Test in the Log
S
Housing Price Equation
We apply the special case of the White test to equation (8.18), where we use the LM form of the
statistic. The important thing to remember is that the chi-square distribution always has two df. The
regression of u^ 2 on lprice, 1 lprice 2 2, where lprice denotes the fitted values from (8.18), produces
R2u^ 2 5 .0392; thus, LM 5 88 1 .0392 2 < 3.45, and the p-value 5 .178. This is stronger evidence of
heteroskedasticity than is provided by the Breusch-Pagan test, but we still fail to reject homoskedasticity at even the 15% level.
Before leaving this section, we should discuss one important caveat. We have interpreted a rejection
using one of the heteroskedasticity tests as evidence of heteroskedasticity. This is appropriate provided we
maintain Assumptions MLR.1 through MLR.4. But, if MLR.4 is violated—in particular, if the functional
form of E 1 y 0 x 2 is misspecified—then a test for heteroskedasticity can reject H0, even if Var 1 y 0 x 2 is constant. For example, if we omit one or more quadratic terms in a regression model or use the level model
when we should use the log, a test for heteroskedasticity can be significant. This has led some economists
to view tests for heteroskedasticity as general misspecification tests. However, there are better, more direct
tests for functional form misspecification, and we will cover some of them in Section 9-1. It is better to
use explicit tests for functional form first, as functional form misspecification is more important than heteroskedasticity. Then, once we are satisfied with the functional form, we can test for heteroskedasticity.
8-4 Weighted Least Squares Estimation
If heteroskedasticity is detected using one of the tests in Section 8-3, we know from Section 8-2 that
one possible response is to use heteroskedasticity-robust statistics after estimation by OLS. Before
the development of heteroskedasticity-robust statistics, the response to a finding of heteroskedasticity
was to specify its form and use a weighted least squares method, which we develop in this section.
As we will argue, if we have correctly specified the form of the variance (as a function of explanatory variables), then weighted least squares (WLS) is more efficient than OLS, and WLS leads to new
t and F statistics that have t and F distributions. We will also discuss the implications of using the
wrong form of the variance in the WLS procedure.
8-4a The Heteroskedasticity Is Known up to a Multiplicative Constant
Let x denote all the explanatory variables in equation (8.10) and assume that
Var 1 u 0 x 2 5 s2h 1 x 2 ,
[8.21]
savi 5 b0 1 biinci 1 ui
[8.22]
Var 1 ui 0 inci 2 5 s inci.
[8.23]
where h 1 x 2 is some function of the explanatory variables that determines the heteroskedasticity.
Because variances must be positive, h 1 x 2 . 0 for all possible values of the independent variables. For
now, we assume that the function h 1 x 2 is known. The population parameter s2 is unknown, but we
will be able to estimate it from a data sample.
For a random drawing from the population, we can write s2i 5 Var 1 ui 0 xi 2 5 s2h 1 xi 2 5 s2hi,
where we again use the notation xi to denote all independent variables for observation i, and hi
changes with each observation because the independent variables change across observations. For
example, consider the simple savings function
2
58860_ch08_hr_262-293.indd 273
10/18/18 4:29 PM
274
PART 1 Regression Analysis with Cross-Sectional Data
Here, h 1 x 2 5 h 1 inc 2 5 inc: the variance of the error is proportional to the level of income. This
means that, as income increases, the variability in savings increases. (If b1 . 0, the expected value of
savings also increases with income.) Because inc is always positive, the variance in equation (8.23) is
always guaranteed to be positive. The standard deviation of ui, conditional on inci, is s !inci.
How can we use the information in equation (8.21) to estimate the bj? Essentially, we take the
original equation,
yi 5 b0 1 b1xi1 1 b2xi2 1 p 1 bkxik 1 ui,
[8.24]
which contains heteroskedastic errors, and transform it into an equation that has homoskedastic errors
(and satisfies the other Gauss-Markov assumptions). Because hi is just a function of xi, ui /!hi has a
zero expected value conditional on xi. Further, because Var 1 ui 0 xi 2 5 E 1 u2i 0 xi 2 5 s2hi, the variance of
ui /!hi (conditional on xi) is s2:
E 3 1 ui / "hi 2 2 4 5 E 1 u2i 2 /hi 5 1 s2hi 2 /hi 5 s2,
where we have suppressed the conditioning on xi for simplicity. We can divide equation (8.24) by !hi
to get
or
yi / "hi 5 b0/ "hi 1 b1 1 xi1/ "hi 2 1 b2 1 xi2/ "hi 2 1 p
1 bk 1 xik/ "hi 2 1 1 ui/ "hi 2
ypi 5 b0xpi0 1 b1xpi1 1 p 1 bkxpik 1 upi ,
[8.25]
[8.26]
where xpi0 5 1/!hi and the other starred variables denote the corresponding original variables divided
by !hi.
Equation (8.26) looks a little peculiar, but the important thing to remember is that we derived it so we
could obtain estimators of the bj that have better efficiency properties than OLS. The intercept b0 in the
original equation (8.24) is now multiplying the variable xpi0 5 1/!hi. Each slope parameter in bj multiplies
a new variable that rarely has a useful interpretation. This should not cause problems if we recall that, for
interpreting the parameters and the model, we always want to return to the original equation (8.24).
In the preceding savings example, the transformed equation looks like
savi/ "inci 5 b0 1 1/ "inci 2 1 b1"inci 1 upi ,
where we use the fact that inci/!inci 5 !inci. Nevertheless, b1 is the marginal propensity to save out
of income, an interpretation we obtain from equation (8.22).
Equation (8.26) is linear in its parameters (so it satisfies MLR.1), and the random sampling
assumption has not changed. Further, upi has a zero mean and a constant variance 1 s2 2 , conditional on xpi .
This means that if the original equation satisfies the first four Gauss-Markov assumptions, then the
transformed equation (8.26) satisfies all five Gauss-Markov assumptions. Also, if ui has a normal
distribution, then upi has a normal distribution with variance s2. Therefore, the transformed equation
satisfies the classical linear model assumptions (MLR.1 through MLR.6) if the original model does so
except for the homoskedasticity assumption.
Because we know that OLS has appealing properties (is BLUE, for example) under the GaussMarkov assumptions, the discussion in the previous paragraph suggests estimating the parameters in
equation (8.26) by ordinary least squares. These estimators, bp0, bp1, . . . , bpk , will be different from the
OLS estimators in the original equation. The bpj are examples of generalized least squares (GLS)
estimators. In this case, the GLS estimators are used to account for heteroskedasticity in the errors.
We will encounter other GLS estimators in Chapter 12.
Because equation (8.26) satisfies all of the ideal assumptions, standard errors, t statistics, and
F statistics can all be obtained from regressions using the transformed variables. The sum of squared
residuals from (8.26) divided by the degrees of freedom is an unbiased estimator of s2. Further, the
GLS estimators, because they are the best linear unbiased estimators of the bj, are necessarily more
58860_ch08_hr_262-293.indd 274
10/18/18 4:29 PM
CHAPTER 8 Heteroskedasticity
275
efficient than the OLS estimators b^ j obtained from the untransformed equation. Essentially, after we
have transformed the variables, we simply use standard OLS analysis. But we must remember to
interpret the estimates in light of the original equation.
The GLS estimators for correcting heteroskedasticity are called weighted least squares (WLS)
estimators. This name comes from the fact that the bpj minimize the weighted sum of squared residuals, where each squared residual is weighted by 1/hi. The idea is that less weight is given to observations with a higher error variance; OLS gives each observation the same weight because it is best
when the error variance is identical for all partitions of the population. Mathematically, the WLS
estimators are the values of the bj that make
p 2 bkxik 2 2/hi
a 1 yi 2 b0 2 b1xi1 2 b2xi2 2
n
[8.27]
i51
as small as possible. Bringing the square root of 1/hi inside the squared residual shows that the weighted
sum of squared residuals is identical to the sum of squared residuals in the transformed variables:
p
p
p
p
p 2 bkxpik 2 2.
a 1 yi 2 b0xi0 2 b1xi1 2 b2xi2 2
n
i51
Because OLS minimizes the sum of squared residuals (regardless of the definitions of the dependent
variable and independent variable), it follows that the WLS estimators that minimize (8.27) are simply the OLS estimators from (8.26). Note carefully that the squared residuals in (8.27) are weighted
by 1/hi, whereas the transformed variables in (8.26) are weighted by 1/!hi.
A weighted least squares estimator can be defined for any set of positive weights. Ordinary least
squares is the special case that gives equal weight to all observations. The efficient procedure, GLS,
weights each squared residual by the inverse of the conditional variance of ui given xi.
Obtaining the transformed variables in equation (8.25) in order to manually perform weighted
least squares can be tedious, and the chance of making mistakes is nontrivial. Fortunately, most modern regression packages have a feature for computing weighted least squares. Typically, along with
the dependent and independent variables in the original model, we just specify the weighting function, 1/hi, appearing in (8.27). That is, we specify weights proportional to the inverse of the variance.
In addition to making mistakes less likely, this forces us to interpret weighted least squares estimates
in the original model. In fact, we can write out the estimated equation in the usual way. The estimates
and standard errors will be different from OLS, but the way we interpret those estimates, standard
errors, and test statistics is the same.
Econometrics packages that have a built-in WLS option will report an R-squared (and adjusted
R-squared) along with WLS estimates and standard errors. Typically, the WLS R-squared is obtained
from the weighted SSR, obtained from minimizing equation (8.27), and a weighted total sum of
squares (SST), obtained by using the same weights but setting all of the slope coefficients in equation
(8.27), b1, b2, . . . , bk, to zero. As a goodness-of-fit measure, this R-squared is not especially useful,
as it effectively measures explained variation in ypi rather than yi. Nevertheless, the WLS R-squareds
computed as just described are appropriate for computing F statistics for exclusion restrictions (provided we have properly specified the variance function). As in the case of OLS, the SST terms cancel,
and so we obtain the F statistic based on the weighted SSR.
The R-squared from running the OLS regression in equation (8.26) is even less useful as a
goodness-of-fit measure, as the computation of SST would make little sense: one would necessarily exclude an intercept from the regression, in which case regression packages typically compute
the SST without properly centering the ypi . This is another reason for using a WLS option that is
pre-programmed in a regression package because at least the reported R-squared properly compares
the model with all of the independent variables to a model with only an intercept. Because the SST
cancels out when testing exclusion restrictions, improperly computing SST does not affect the
R-squared form of the F statistic. Nevertheless, computing such an R-squared tempts one to think the
equation fits better than it does.
58860_ch08_hr_262-293.indd 275
10/18/18 4:29 PM
276
PART 1 Regression Analysis with Cross-Sectional Data
Example 8.6
Financial Wealth Equation
We now estimate equations that explain net total financial wealth (nettfa, measured in $1,000s) in
terms of income (inc, also measured in $1,000s) and some other variables, including age, gender,
and an indicator for whether the person is eligible for a 401(k) pension plan. We use the data on
single people 1 fsize 5 1 2 in 401KSUBS. In Computer Exercise C12 in Chapter 6, it was found that
a specific quadratic function in age, namely 1 age 2 25 2 2, fit the data just as well as an unrestricted
quadratic. Plus, the restricted form gives a simplified interpretation because the minimum age in the
sample is 25: nettfa is an increasing function of age after age 5 25.
The results are reported in Table 8.1. Because we suspect heteroskedasticity, we report the
­heteroskedasticity-robust standard errors for OLS. The weighted least squares estimates, and their
standard errors, are obtained under the assumption Var 1 u 0 inc 2 5 s2inc.
Without controlling for other factors, another dollar of income is estimated to increase nettfa by
about 82¢ when OLS is used; the WLS estimate is smaller, about 79¢. The difference is not large; we
certainly do not expect them to be identical. The WLS coefficient does have a smaller standard error
than OLS, almost 40% smaller, provided we assume the model Var 1 nettfa 0 inc 2 5 s2inc is correct.
Adding the other controls reduced the inc coefficient somewhat, with the OLS estimate still
larger than the WLS estimate. Again, the WLS estimate of binc is more precise. Age has an increasing
effect starting at age 5 25, with the OLS estimate showing a larger effect. The WLS estimate of bage
is more precise in this case. Gender does not have a statistically significant effect on nettfa, but being
eligible for a 401(k) plan does: the OLS estimate is that those eligible, holding fixed income, age, and
gender, have net total financial assets about $6,890 higher. The WLS estimate is substantially below
the OLS estimate and suggests a misspecification of the functional form in the mean equation. (One
possibility is to interact e401k and inc; see Computer
Exercise C11.)
G o i n g F u rt h e r 8 . 3
Using WLS, the F statistic for joint significance
Using the OLS residuals obtained from the
of 1 age 2 25 2 2, male, and e401k is about 30.8 if we
OLS regression reported in column (1) of
use the R-squareds reported in Table 8.1. With 3 and
Table 8.1, the regression of u^ 2 on inc yields
2,012 degrees of freedom, the p-value is zero to more
a t statistic of 2.96. Does it appear we
than 15 decimal places; of course, this is not surprisshould worry about heteroskedasticity in the
ing given the very large t statistics for the age and
financial wealth equation?
401(k) variables.
Table 8.1 Dependent Variable: nettfa
Independent
Variables
(1)
OLS
(2)
WLS
(3)
OLS
(4)
WLS
inc
.821
(.104)
.787
(.063)
.771
(.100)
.740
(.064)
1 age 2 25 2 2
—
—
.0251
(.0043)
.0175
(.0019)
male
—
—
2.48
(2.06)
1.84
(1.56)
e401k
—
—
6.89
(2.29)
5.19
(1.70)
210.57
(2.53)
29.58
(1.65)
220.98
(3.50)
216.70
(1.96)
Observations
2,017
2,017
2,017
2,017
R-squared
.0827
.0709
.1279
.1115
intercept
58860_ch08_hr_262-293.indd 276
10/18/18 4:29 PM
CHAPTER 8 Heteroskedasticity
277
Assuming that the error variance in the financial wealth equation has a variance proportional to
income is essentially arbitrary. In fact, in most cases, our choice of weights in WLS has a degree of
arbitrariness. However, there is one case in which the weights needed for WLS arise naturally from
an underlying econometric model. This happens when, instead of using individual-level data, we only
have averages of data across some group or geographic region. For example, suppose we are interested in determining the relationship between the amount a worker contributes to his or her 401(k)
pension plan as a function of the plan generosity. Let i denote a particular firm and let e denote an
employee within the firm. A simple model is
contribi,e 5 b0 1 b1earnsi,e 1 b2agei,e 1 b3mratei 1 ui,e,
[8.28]
where contribi,e is the annual contribution by employee e who works for firm i, earnsi,e is annual earnings for this person, and agei,e is the person’s age. The variable mratei is the amount the firm puts into
an employee’s account for every dollar the employee contributes.
If (8.28) satisfies the Gauss-Markov assumptions, then we could estimate it, given a sample of
individuals across various employers. Suppose, however, that we only have average values of contributions, earnings, and age by employer. In other words, individual-level data are not available. Thus,
let contribi denote average contribution for people at firm i, and similarly for earnsi and agei. Let mi
denote the number of employees at firm i; we assume that this is a known quantity. Then, if we average equation (8.28) across all employees at firm i, we obtain the firm-level equation
contribi 5 b0 1 b1earnsi 1 b2agei 1 b3mratei 1 ui,
[8.29]
where ui 5 m21
i g e51 ui,e is the average error across all employees in firm i. If we have n firms in our
sample, then (8.29) is just a standard multiple linear regression model that can be estimated by OLS.
The estimators are unbiased if the original model (8.28) satisfies the Gauss-Markov assumptions and
the individual errors ui,e are independent of the firm’s size, mi [because then the expected value of ui,
given the explanatory variables in (8.29), is zero].
If the individual-level equation (8.28) satisfies the homoskedasticity assumption, and the ­errors
within firm i are uncorrelated across employees, then we can show that the firm-level equation
(8.29) has a particular kind of heteroskedasticity. Specifically, if Var 1 ui,e 2 5 s2 for all i and e, and
Cov 1 ui,e, ui,g 2 5 0 for every pair of employees e 2 g within firm i, then Var 1 ui 2 5 s2/mi; this is just
the usual formula for the variance of an average of uncorrelated random variables with common variance. In other words, the variance of the error term ui decreases with firm size. In this case, hi 5 1/mi,
and so the most efficient procedure is weighted least squares, with weights equal to the number of
employees at the firm 1 1/hi 5 mi 2 . This ensures that larger firms receive more weight. This gives us
an efficient way of estimating the parameters in the individual-level model when we only have averages at the firm level.
A similar weighting arises when we are using per capita data at the city, county, state, or country level. If the individual-level equation satisfies the Gauss-Markov assumptions, then the error in
the per capita equation has a variance proportional to one over the size of the population. Therefore,
weighted least squares with weights equal to the population is appropriate. For example, suppose we
have city-level data on per capita beer consumption (in ounces), the percentage of people in the population over 21 years old, average adult education levels, average income levels, and the city price of
beer. Then, the city-level model
mi
beerpc 5 b0 1 b1perc21 1 b2avgeduc 1 b3incpc 1 b4 price 1 u
can be estimated by weighted least squares, with the weights being the city population.
The advantage of weighting by firm size, city population, and so on relies on the underlying
individual equation being homoskedastic. If heteroskedasticity exists at the individual level, then the
proper weighting depends on the form of heteroskedasticity. Further, if there is correlation across
­errors within a group (say, firm), then Var 1 ui 2 2 s2/mi; see Problem 7. Uncertainty about the form of
Var 1 ui 2 in equations such as (8.29) is why more and more researchers simply use OLS and compute
58860_ch08_hr_262-293.indd 277
10/18/18 4:29 PM
278
PART 1 Regression Analysis with Cross-Sectional Data
robust standard errors and test statistics when estimating models using per capita data. An alternative is to weight by group size but to report the heteroskedasticity-robust statistics in the WLS estimation. This ensures that, while the estimation is efficient if the individual-level model satisfies the
Gauss-Markov assumptions, heteroskedasticity at the individual level or within-group correlation are
accounted for through robust inference.
8-4b The Heteroskedasticity Function
Must Be Estimated: Feasible GLS
In the previous subsection, we saw some examples of where the heteroskedasticity is known up to a
multiplicative form. In most cases, the exact form of heteroskedasticity is not obvious. In other words, it
is difficult to find the function h 1 xi 2 of the previous section. Nevertheless, in many cases we can model
the function h and use the data to estimate the unknown parameters in this model. This results in an estimate of each hi, denoted as h^ i. Using h^ i instead of hi in the GLS transformation yields an estimator called
the feasible GLS (FGLS) estimator. Feasible GLS is sometimes called estimated GLS, or EGLS.
There are many ways to model heteroskedasticity, but we will study one particular, fairly flexible
approach. Assume that
Var 1 u 0 x 2 5 s2exp 1 d0 1 d1x1 1 d2x2 1 p 1 dkxk 2 ,
[8.30]
where x1, x2, . . . , xk are the independent variables appearing in the regression model [see equation
(8.1)], and the dj are unknown parameters. Other functions of the xj can appear, but we will focus primarily on (8.30). In the notation of the previous subsection, h 1 x 2 5 exp 1 d0 1 d1x1 1 d2x2 1 p 1 dkxk 2 .
You may wonder why we have used the exponential function in (8.30). After all, when testing
for heteroskedasticity using the Breusch-Pagan test, we assumed that heteroskedasticity was a linear
function of the xj. Linear alternatives such as (8.12) are fine when testing for heteroskedasticity, but
they can be problematic when correcting for heteroskedasticity using weighted least squares. We have
encountered the reason for this problem before: linear models do not ensure that predicted values are
positive, and our estimated variances must be positive in order to perform WLS.
If the parameters dj were known, then we would just apply WLS, as in the previous subsection.
This is not very realistic. It is better to use the data to estimate these parameters, and then to use these
estimates to construct weights. How can we estimate the dj? Essentially, we will transform this equation into a linear form that, with slight modification, can be estimated by OLS.
Under assumption (8.30), we can write
u2 5 s2exp 1 d0 1 d1x1 1 d2x2 1 p 1 dkxk 2 v,
where v has a mean equal to unity, conditional on x 5 (x1, x2, . . . , xk). If we assume that v is actually
independent of x, we can write
log 1 u2 2 5 a0 1 d1x1 1 d2x2 1 p 1 dkxk 1 e,
[8.31]
where e has a zero mean and is independent of x; the intercept in this equation is different from d0, but this
is not important in implementing WLS. The dependent variable is the log of the squared e­ rror. Because
(8.31) satisfies the Gauss-Markov assumptions, we can get unbiased estimators of the dj by ­using OLS.
As usual, we must replace the unobserved u with the OLS residuals. Therefore, we run the
­regression of
log 1 u^ 2 2 on x1, x2, . . . , xk.
[8.32]
h^ i 5 exp 1 g^ i 2 .
[8.33]
Actually, what we need from this regression are the fitted values; call these g^ i. Then, the estimates of
hi are simply
We now use WLS with weights 1/h^ i in place of 1/hi in equation (8.27). We summarize the steps.
58860_ch08_hr_262-293.indd 278
10/18/18 4:29 PM
CHAPTER 8 Heteroskedasticity
279
A Feasible GLS Procedure to Correct for Heteroskedasticity:
1.
2.
3.
4.
5.
Run the regression of y on x1, x2, . . . , xk and obtain the residuals, u^ .
Create log 1 u^ 2 2 by first squaring the OLS residuals and then taking the natural log.
Run the regression in equation (8.32) and obtain the fitted values, g^ .
Exponentiate the fitted values from (8.32): h^ 5 exp 1 g^ 2 .
Estimate the equation
y 5 b0 1 b1x1 1 p 1 bkxk 1 u
by WLS, using weights 1/h^ . In other words, we replace hi with h^ i in equation (8.27). Remember,
the squared residual for observation i gets weighted by 1/h^ i. If instead we first transform all variables and run OLS, each variable gets multiplied by 1/"h^ i, including the intercept.
If we could use hi rather than h^ i in the WLS procedure, we know that our estimators would be
unbiased; in fact, they would be the best linear unbiased estimators, assuming that we have properly
modeled the heteroskedasticity. Having to estimate hi using the same data means that the FGLS estimator is no longer unbiased (so it cannot be BLUE, either). Nevertheless, the FGLS estimator is consistent and asymptotically more efficient than OLS. This is difficult to show because of estimation of
the variance parameters. But if we ignore this—as it turns out we may—the proof is similar to showing that OLS is efficient in the class of estimators in Theorem 5.3. At any rate, for large sample sizes,
FGLS is an attractive alternative to OLS when there is evidence of heteroskedasticity that inflates the
standard errors of the OLS estimates.
We must remember that the FGLS estimators are estimators of the parameters in the usual population model
y 5 b0 1 b1x1 1 p 1 bkxk 1 u.
Just as the OLS estimates measure the marginal impact of each xj on y, so do the FGLS estimates. We
use the FGLS estimates in place of the OLS estimates because the FGLS estimators are more efficient
and have associated test statistics with the usual t and F distributions, at least in large samples. If we
have some doubt about the variance specified in equation (8.30), we can use heteroskedasticity-robust
standard errors and test statistics in the transformed equation.
Another useful alternative for estimating hi is to replace the independent variables in regression
(8.32) with the OLS fitted values and their squares. In other words, obtain the g^ i as the fitted values
from the regression of
log 1 u^ 2 2 on y^ , y^ 2
[8.34]
and then obtain the h^ i exactly as in equation (8.33). This changes only step (3) in the previous
procedure.
If we use regression (8.32) to estimate the variance function, you may be wondering if we can
simply test for heteroskedasticity using this same regression (an F or LM test can be used). In fact,
Park (1966) suggested this. Unfortunately, when compared with the tests discussed in Section 8-3,
the Park test has some problems. First, the null hypothesis must be something stronger than homoskedasticity: effectively, u and x must be independent. This is not required in the Breusch-Pagan or
White tests. Second, using the OLS residuals u^ in place of u in (8.32) can cause the F statistic to
deviate from the F distribution, even in large sample sizes. This is not an issue in the other tests we
have covered. For these reasons, the Park test is not recommended when testing for heteroskedasticity.
Regression (8.32) works well for weighted least squares because we only need consistent estimators
of the dj, and regression (8.32) certainly delivers those.
58860_ch08_hr_262-293.indd 279
10/18/18 4:29 PM
280
PART 1 Regression Analysis with Cross-Sectional Data
Demand for Cigarettes
Example 8.7
We use the data in SMOKE to estimate a demand function for daily cigarette consumption. Because
most people do not smoke, the dependent variable, cigs, is zero for most observations. A linear model
is not ideal because it can result in negative predicted values. Nevertheless, we can still learn something about the determinants of cigarette smoking by using a linear model.
The equation estimated by ordinary least squares, with the usual OLS standard errors in parentheses, is
cigs 5 23.64 1 .880 log 1 income 2 2 .751 log 1 cigpric 2
1 24.08 2 1 .728 2
1 5.773 2
2.501 educ 1 .771 age 2 .0090 age2 2 2.83 restaurn
1 .167 2
1 .160 2
1 .0017 2
1 1.11 2
[8.35]
n 5 807, R2 5 .0526,
where
cigs 5 number of cigarettes smoked per day.
income 5 annual income.
cigpric 5 the per-pack price of cigarettes (in cents).
educ 5 years of schooling.
age 5 age measured in years.
restaurn 5 a binary indicator equal to unity if the person resides in a state with restaurant
smoking restrictions.
Because we are also going to do weighted least squares, we do not report the heteroskedasticityrobust standard errors for OLS. (Incidentally, 13 out of the 807 fitted values are less than zero; this is
less than 2% of the sample and is not a major cause for concern.)
Neither income nor cigarette price is statistically significant in (8.35), and their effects are
not practically large. For example, if income increases by 10%, cigs is predicted to increase by
1 .880/100 2 1 10 2 5 .088, or less than one-tenth of a cigarette per day. The magnitude of the price
effect is similar.
Each year of education reduces the average cigarettes smoked per day by one-half of a cigarette,
and the effect is statistically significant. Cigarette smoking is also related to age, in a quadratic fashion. Smoking increases with age up until age 5 .771/ 3 2 1 .009 2 4 < 42.83, and then smoking decreases
with age. Both terms in the quadratic are statistically significant. The presence of a restriction on
smoking in restaurants decreases cigarette smoking by almost three cigarettes per day, on average.
Do the errors underlying equation (8.35) contain heteroskedasticity? The Breusch-Pagan regression of the squared OLS residuals on the independent variables in (8.35) [see equation (8.14)] produces R2u^ 2 5 .040. This small R-squared may seem to indicate no heteroskedasticity, but we must
remember to compute either the F or LM statistic. If the sample size is large, a seemingly small R2u^ 2 can
result in a very strong rejection of homoskedasticity. The LM statistic is LM 5 807 1 .040 2 5 32.28,
and this is the outcome of a x26 random variable. The p-value is less than .000015, which is very
strong evidence of heteroskedasticity.
Therefore, we estimate the equation using the feasible GLS procedure based on equation (8.32).
The weighted least squares estimates are
cigs 5 5.64 1 1.30 log 1 income 2 2 2.94 log 1 cigpric 2
1 17.80 2 1 .44 2
1 4.46 2
2.463 educ 1 .482 age 2 .0056 age2 2 3.46 restaurn
1 .120 2
1 .097 2
1 .0009 2
1 .80 2
[8.36]
n 5 807, R2 5 .1134.
58860_ch08_hr_262-293.indd 280
10/18/18 4:29 PM
CHAPTER 8 Heteroskedasticity
281
The income effect is now statistically significant and larger in magnitude. The price effect is also
notably bigger, but it is still statistically insignificant. [One reason for this is that cigpric varies only
across states in the sample, and so there is much less variation in log(cigpric) than in log(income),
educ, and age.]
The estimates on the other variables have, naturally, changed somewhat, but the basic story is still
the same. Cigarette smoking is negatively related to schooling, has a quadratic relationship with age,
and is negatively affected by restaurant smoking restrictions.
We must be a little careful in computing F statistics for testing multiple hypotheses after estimation by WLS. (This is true whether the sum of squared residuals or R-squared form of the F statistic
is used.) It is important that the same weights be used to estimate the unrestricted and restricted
models. We should first estimate the unrestricted model by OLS. Once we have obtained the weights,
we can use them to estimate the restricted model as well. The F statistic can be computed as usual.
Fortunately, many regression packages have a simple command for testing joint restrictions after
WLS estimation, so we need not perform the restricted regression ourselves.
Example 8.7 hints at an issue that sometimes arises in applications of weighted least squares: the
OLS and WLS estimates can be substantially different. This is not such a big problem in the demand
for cigarettes equation because all the coefficients maintain the same signs, and the biggest changes
are on variables that were statistically insignificant
when the equation was estimated by OLS. The OLS
G o i n g F u rt h e r 8 . 4
and WLS estimates will always differ due to samLet u^ i be the WLS residuals from (8.36),
pling error. The issue is whether their difference is
which are not weighted, and let cigsi be
enough to change important conclusions.
the fitted values. (These are obtained
If OLS and WLS produce statistically significant
using the same formulas as OLS; they difestimates that differ in sign—for example, the OLS
fer because of different estimates of the
price elasticity is positive and significant, while the
bj.) One way to determine whether heteroWLS price elasticity is negative and s­ ignificant—
skedasticity has been eliminated is to use
or the difference in magnitudes of the estimates is
the u^ 2i /h^ i 5 1 u^ i/"h^ i 2 2 in a test for heteroskepractically large, we should be suspicious. Typically,
dasticity. [If hi 5 Var 1 ui 0 xi 2 , then the transthis indicates that one of the other Gauss-Markov
formed residuals should have little evidence
assumptions is false, particularly the zero condiof heteroskedasticity.] There are many postional mean assumption on the error (MLR.4). If
sibilities, but one—based on White’s test
in the transformed equation—is to regress
E 1 y 0 x 2 2 b0 1 b1x1 1 p 1 bkxk , then OLS and
u^ 2i /h^ i on cigsi / "h^ i and cigs2i /h^ i (including an
WLS have different expected values and probability
intercept). The joint F statistic when we use
limits. For WLS to be consistent for the bj, it is not
SMOKE is 11.15. Does it appear that our
enough for u to be uncorrelated with each xj; we need
correction for heteroskedasticity has actually
the stronger assumption MLR.4 in the linear model
eliminated the heteroskedasticity?
MLR.1. Therefore, a significant difference between
OLS and WLS can indicate a functional form misspecification in E 1 y 0 x 2 . The Hausman test [Hausman (1978)] can be used to formally compare the
OLS and WLS estimates to see if they differ by more than sampling error suggests they should, but
this test is beyond the scope of this text. In many cases, an informal “eyeballing” of the estimates is
sufficient to detect a problem.
8-4c What If the Assumed Heteroskedasticity Function Is Wrong?
We just noted that if OLS and WLS produce very different estimates, it is likely that the conditional
mean E 1 y 0 x 2 is misspecified. What are the properties of WLS if the variance function we use is misspecified in the sense that Var 1 y 0 x 2 2 s2h 1 x 2 for our chosen function h(x)? The most important issue
58860_ch08_hr_262-293.indd 281
10/18/18 4:29 PM
282
PART 1 Regression Analysis with Cross-Sectional Data
is whether misspecification of h(x) causes bias or inconsistency in the WLS estimator. Fortunately, the
answer is no, at least under MLR.4. Recall that, if E 1 u 0 x 2 5 0, then any function of x is uncorrelated
with u, and so the weighted error, u/ "h 1 x 2 , is uncorrelated with the weighted regressors, xj / "h 1 x 2 ,
for any function h(x) that is always positive. This is why, as we just discussed, we can take large differences between the OLS and WLS estimators as indicative of functional form misspecification. If
we estimate parameters in the function, say h 1 x, d^ 2 , then we can no longer claim that WLS is unbiased, but it will generally be consistent (whether or not the variance function is correctly specified).
If WLS is at least consistent under MLR.1 to MLR.4, what are the consequences of using WLS
with a misspecified variance function? There are two. The first, which is very important, is that the
usual WLS standard errors and test statistics, computed under the assumption that Var 1 y 0 x 2 5 s2h 1 x 2 ,
are no longer valid, even in large samples. For example, the WLS estimates and standard errors in column (4)
of Table 8.1 assume that Var 1 nettfa 0 inc, age, male, e401k 2 5 Var 1 nettfa 0 inc 2 5 s2inc; so we are
assuming not only that the variance depends just on income, but also that it is a linear function of income.
If this assumption is false, the standard errors (and any statistics we obtain using those standard errors) are
not valid. Fortunately, there is an easy fix: just as we can obtain standard errors for the OLS estimates that
are robust to arbitrary heteroskedasticity, we can obtain standard errors for WLS that allow the variance
function to be arbitrarily misspecified. It is easy to see why this works. Write the transformed equation as
yi / "hi 5 b0 1 1 / "hi 2 1 b1 1 xi1 / "hi 2 1 p 1 bk 1 xik / "hi 2 1 ui / "hi.
Now, if Var 1 ui 0 xi 2 2 s2hi, then the weighted error ui / "hi is heteroskedastic. So we can just apply
the usual heteroskedasticity-robust standard errors after estimating this equation by OLS—which,
remember, is identical to WLS.
To see how robust inference with WLS works in practice, column (1) of Table 8.2 reproduces the
last column of Table 8.1, and column (2) contains standard errors robust to Var 1 ui 0 xi 2 2 s2inci.
The standard errors in column (2) allow the variance function to be misspecified. We see that, for
the income and age variables, the robust standard errors are somewhat above the usual WLS standard
errors—certainly by enough to stretch the confidence intervals. On the other hand, the robust standard
errors for male and e401k are actually smaller than those that assume a correct variance function. We
saw this could happen with the heteroskedasticity-robust standard errors for OLS, too.
Even if we use flexible forms of variance functions, such as that in (8.30), there is no guarantee
that we have the correct model. While exponential heteroskedasticity is appealing and reasonably flexible, it is, after all, just a model. Therefore, it is always a good idea to compute fully robust standard
errors and test statistics after WLS estimation.
Table 8.2 WLS Estimation of the nettfa Equation
Independent
Variables
With Nonrobust
Standard Errors
With Robust
Standard Errors
inc
.740
(.064)
.740
(.075)
(age – 25)2
.0175
(.0019)
.0175
(.0026)
male
1.84
(1.56)
1.84
(1.31)
e401k
5.19
(1.70)
5.19
(1.57)
216.70
(1.96)
216.70
(2.24)
Observations
2,017
2,017
R-squared
.1115
.1115
intercept
58860_ch08_hr_262-293.indd 282
10/18/18 4:29 PM
CHAPTER 8 Heteroskedasticity
283
A modern criticism of WLS is that if the variance function is misspecified, it is not guaranteed
to be more efficient than OLS. In fact, that is the case: if Var 1 y 0 x 2 is neither constant nor equal to
s2h 1 x 2 , where h(x) is the proposed model of heteroskedasticity, then we cannot rank OLS and WLS
in terms of variances (or asymptotic variances when the variance parameters must be estimated).
However, this theoretically correct criticism misses an important practical point. Namely, in cases of
strong heteroskedasticity, it is often better to use a wrong form of heteroskedasticity and apply WLS
than to ignore heteroskedasticity altogether in estimation and use OLS. Models such as (8.30) can
well approximate a variety of heteroskedasticity functions and may produce estimators with smaller
(asymptotic) variances. Even in Example 8.6, where the form of heteroskedasticity was assumed to
have the simple form Var 1 nettfa 0 x 2 5 s2inc, the fully robust standard errors for WLS are well below
the fully robust standard errors for OLS. (Comparing robust standard errors for the two estimators
puts them on equal footing: we assume neither homoskedasticity nor that the variance has the form
s2inc.) For example, the robust standard error for the WLS estimator of binc is about .075, which is
25% lower than the robust standard error for OLS (about .100). For the coefficient on 1 age 2 25 2 2,
the robust standard error of WLS is about .0026, almost 40% below the robust standard error for OLS
(about .0043).
8-4d Prediction and Prediction Intervals with Heteroskedasticity
If we start with the standard linear model under MLR.1 to MLR.4, but allow for heteroskedasticity
of the form Var 1 y 0 x 2 5 s2h 1 x 2 [see equation (8.21)], the presence of heteroskedasticity affects the
point prediction of y only insofar as it affects estimation of the bj. Of course, it is natural to use WLS
on a sample of size n to obtain the b^ j. Our prediction of an unobserved outcome, y0, given known
^ . This
values of the explanatory variables x0, has the same form as in Section 6-4: y^ 0 5 b^ 0 1 x0b
makes sense: once we know E 1 y 0 x 2 , we base our prediction on it; the structure of Var 1 y 0 x 2 plays no
direct role.
On the other hand, prediction intervals do depend directly on the nature of Var 1 y 0 x 2 . Recall in
Section 6-4 that we constructed a prediction interval under the classical linear model assumptions.
Suppose now that all the CLM assumptions hold except that (8.21) replaces the homoskedasticity
assumption, MLR.5. We know that the WLS estimators are BLUE and, because of normality, have (conditional) normal distributions. We can obtain se 1 y^ 0 2 using the same method in Section 6-4, except that
now we use WLS. [A simple approach is to write yi 5 u 0 1 b1 1 xi1 2 x01 2 1 p 1 bk 1 xik 2 x0k 2 1 ui,
where the x0j are the values of the explanatory variables for which we want a predicted value of y.
We can estimate this equation by WLS and then obtain y^ 0 5 u^ 0 and se 1 y^ 0 2 5 se 1 u^ 0 2 .] We also need
to estimate the standard deviation of u0, the unobserved part of y0. But Var 1 u0 0 x 5 x0 2 5 s2h 1 x0 2 ,
and so se 1 u0 2 5 s^ "h 1 x0 2 , where s^ is the standard error of the regression from the WLS estimation.
Therefore, a 95% prediction interval is
y^ 0 6 t.025 ? se 1 e^ 0 2 ,
[8.37]
log 1 y 2 5 b0 1 b1x1 1 p 1 bkxk 1 u,
[8.38]
where se 1 e^ 2 5 5 3 se 1 y^ 2 4 1 s^ h 1 x 2 6 .
This interval is exact only if we do not have to estimate the variance function. If we estimate
parameters, as in model (8.30), then we cannot obtain an exact interval. In fact, accounting for
the estimation error in the b^ j and the d^ j (the variance parameters) becomes very difficult. We saw
two examples in Section 6-4 where the estimation error in the parameters was swamped by the
variation in the unobservables, u0. Therefore, we might still use equation (8.37) with h 1 x0 2 simply
replaced by h^ 1 x0 2 . In fact, if we are to ignore the parameter estimation error entirely, we can drop
se 1 y^ 0 2 from se 1 e^ 0 2 . [Remember, se 1 y^ 0 2 converges to zero at the rate 1/!n, while se 1 u^ 0 2 is roughly
constant.]
We can also obtain a prediction for y in the model
0
58860_ch08_hr_262-293.indd 283
0
2
2
0
1/2
10/18/18 4:29 PM
284
PART 1 Regression Analysis with Cross-Sectional Data
where u is heteroskedastic. We assume that u has a conditional normal distribution with a specific form of
heteroskedasticity. We assume the exponential form in equation (8.30), but add the normality assumption:
u 0 x1, x2, p , xk , Normal 3 0, exp 1 d0 1 d1x1 1 p 1 dkxk 2 4 .
[8.39]
E 1 y 0 x 2 5 exp 1 b0 1 xb 1 exp 1 d0 1 xd 2 /2 2 .
[8.40]
g^ i 5 a^ 0 1 d^ 1xi1 1 p 1 d^ kxik,
[8.41]
As a notational shorthand, write the variance function as exp 1 d0 1 xd 2 . Then, because log(y) given x
has a normal distribution with mean b0 1 xb and variance exp 1 d0 1 xd 2 , it follows that
Now, we estimate the bj and dj using WLS estimation of (8.38). That is, after using OLS to obtain the
residuals, run the regression in (8.32) to obtain fitted values,
and then compute the h^ i as in (8.33). Using these h^ i, obtain the WLS estimates, b^ j, and also compute s^ 2 from the weighted squared residuals. Now, compared with the original model for
Var 1 u 0 x 2 , d0 5 a0 1 log 1 s2 2 , and so Var 1 u 0 x 2 5 s2 exp 1 a0 1 d1x1 1 p 1 dkxk 2 . Therefore, the
­estimated variance is s^ 2 exp 1 g^ i 2 5 s^ 2h^ i, and the fitted value for yi is
y^ i 5 exp 1 logyi 1 s^ 2h^ i / 2 2 .
[8.42]
E^ 1 y 0 x 5 x0 2 5 exp 1 b^ 0 1 x0b^ 1 s^ 2 exp 1 a^ 0 1 x0d^ 2 /2 2 ,
[8.43]
We can use these fitted values to obtain an R-squared measure, as described in Section 6-4: use the
squared correlation coefficient between yi and y^ i.
For any values of the explanatory variables x0, we can estimate E 1 y 0 x 5 x0 2 as
where
b^ j 5 the WLS estimates.
a^ 0 5 the intercept in (8.41).
d^ j 5 the slopes from the same regression.
s^ 2 is obtained from the WLS estimation.
Obtaining a proper standard error for the prediction in (8.42) is very complicated analytically, but, as
in Section 6-4, it would be fairly easy to obtain a standard error using a resampling method such as
the bootstrap described in Appendix 6A.
Obtaining a prediction interval is more of a challenge when we estimate a model for heteroskedasticity, and a full treatment is complicated. Nevertheless, we saw in Section 6-4 two examples where
the error variance swamps the estimation error, and we would make only a small mistake by ­ignoring
the estimation error in all parameters. Using arguments similar to those in Section 6-4, an ­approximate
95% prediction interval (for large sample sizes) is exp 3 21.96 ? s^ "h^ 1 x0 2 4 exp 1 b^ 0 1 x0b^ 2 to
exp 3 1.96 ? s^ "h^ 1 x0 2 4 exp 1 b^ 0 1 x0b^ 2 , where h^ 1 x0 2 is the estimated variance function evaluated at
x0, h^ 1 x0 2 5 exp 1 a^ 0 1 d^ 1x01 1 p 1 d^ kx0k 2 . As in Section 6-4, we obtain this approximate interval by
simply exponentiating the endpoints.
8-5 The Linear Probability Model Revisited
As we saw in Section 7-5, when the dependent variable y is a binary variable, the model must contain
heteroskedasticity, unless all of the slope parameters are zero. We are now in a position to deal with
this problem.
The simplest way to deal with heteroskedasticity in the linear probability model is to continue to
use OLS estimation, but to also compute robust standard errors in test statistics. This ignores the fact
that we actually know the form of heteroskedasticity for the LPM. Nevertheless, OLS estimation of
the LPM is simple and often produces satisfactory results.
58860_ch08_hr_262-293.indd 284
10/18/18 4:29 PM
CHAPTER 8 Heteroskedasticity
Example 8.8
285
Labor Force Participation of Married Women
In the labor force participation example in Section 7-5 [see equation (7.29)], we reported the usual
OLS standard errors. Now, we compute the heteroskedasticity-robust standard errors as well. These
are reported in brackets below the usual standard errors:
inlf 5 .586 2 .0034 nwifeinc 1 .038 educ 1 .039 exper
1 .154 2 1 .0014 2
1 .007 2
1 .006 2
3 .151 4 3 .0015 4
3 .007 4
3 .006 4
2.00060 exper2 2 .016 age 2 .262 kidslt6 1 .0130 kidsge6
1 .00018 2
1 .002 2
1 .034 2
1 .0132 2
3 .00019 4
3 .002 4
3 .032 4
3 .0135 4
[8.44]
n 5 753, R2 5 .264.
Several of the robust and OLS standard errors are the same to the reported degree of precision; in all
cases, the differences are practically very small. Therefore, while heteroskedasticity is a problem in
theory, it is not in practice, at least not for this example. It often turns out that the usual OLS standard
errors and test statistics are similar to their heteroskedasticity-robust counterparts. Furthermore, it
requires a minimal effort to compute both.
Generally, the OLS estimators are inefficient in the LPM. Recall that the conditional variance of
y in the LPM is
where
Var 1 y 0 x 2 5 p 1 x 2 3 1 2 p 1 x 2 4 ,
[8.45]
p 1 x 2 5 b0 1 b1x1 1 p 1 bkxk
[8.46]
h^ i 5 y^ i 1 1 2 y^ i 2 ,
[8.47]
is the response probability (probability of success, y 5 1). It seems natural to use weighted least
squares, but there are a couple of hitches. The probability p(x) clearly depends on the unknown population parameters, bj. Nevertheless, we do have unbiased estimators of these parameters, namely the
OLS estimators. When the OLS estimators are plugged into equation (8.46), we obtain the OLS fitted
values. Thus, for each observation i, Var 1 yi 0 xi 2 is estimated by
where y^ i is the OLS fitted value for observation i. Now, we apply feasible GLS, just as in Section 8-4.
Unfortunately, being able to estimate hi for each i does not mean that we can proceed directly
with WLS estimation. The problem is one that we briefly discussed in Section 7-5: the fitted values
y^ i need not fall in the unit interval. If either y^ i , 0 or y^ i . 1, equation (8.47) shows that h^ i will be
negative. Because WLS proceeds by multiplying observation i by 1 / "h^ i, the method will fail if h^ i is
negative (or zero) for any observation. In other words, all of the weights for WLS must be positive.
In some cases, 0 , y^ i , 1 for all i, in which case WLS can be used to estimate the LPM. In
cases with many observations and small probabilities of success or failure, it is very common to find
some fitted values outside the unit interval. If this happens, as it does in the labor force participation
example in equation (8.44), it is easiest to abandon WLS and to report the heteroskedasticity-robust
statistics. An alternative is to adjust those fitted values that are less than zero or greater than unity, and
then to apply WLS. One suggestion is to set y^ i 5 .01 if y^ i , 0 and y^ i 5 .99 if y^ i . 1. Unfortunately,
this requires an arbitrary choice on the part of the researcher—for example, why not use .001 and .999
as the adjusted values? If many fitted values are outside the unit interval, the adjustment to the fitted
values can affect the results; in this situation, it is probably best to just use OLS.
58860_ch08_hr_262-293.indd 285
10/18/18 4:29 PM
286
PART 1 Regression Analysis with Cross-Sectional Data
Estimating the Linear Probability Model by Weighted Least Squares:
1. Estimate the model by OLS and obtain the fitted values, y^ .
2. Determine whether all of the fitted values are inside the unit interval. If so, proceed to step (3).
If not, some adjustment is needed to bring all fitted values into the unit interval.
3. Construct the estimated variances in equation (8.47).
4. Estimate the equation
y 5 b0 1 b1x1 1 p 1 bkxk 1 u
by WLS, using weights 1/h^ .
Example 8.9
Determinants of Personal Computer Ownership
We use the data in GPA1 to estimate the probability of owning a computer. Let PC denote a binary
indicator equal to unity if the student owns a computer, and zero otherwise. The variable hsGPA is
high school GPA, ACT is achievement test score, and parcoll is a binary indicator equal to unity if
at least one parent attended college. (Separate college indicators for the mother and the father do not
yield individually significant results, as these are pretty highly correlated.)
The equation estimated by OLS is
PC 5 2.0004 1 .065 hsGPA 1 .0006 ACT 1 .221 parcoll
1 .4905 2 1 .137 2
1 .0155 2
1 .093 2
3 .4888 4 3 .139 4
3 .0158 4
3 .087 4
2
n 5 141, R 5 .0415.
[8.48]
Just as with Example 8.8, there are no striking differences between the usual and robust standard
errors. Nevertheless, we also estimate the model by WLS. Because all of the OLS fitted values are
inside the unit interval, no adjustments are needed:
PC 5 .026 1 .033 hsGPA 1 .0043 ACT 1 .215 parcoll
1 .477 2 1 .130 2
1 .0155 2
1 .086 2
2
n 5 142, R 5 .0464.
[8.49]
There are no important differences in the OLS and WLS estimates. The only significant explanatory
variable is parcoll, and in both cases we estimate that the probability of PC ownership is about .22
higher if at least one parent attended college.
Summary
We began by reviewing the properties of ordinary least squares in the presence of heteroskedasticity.
Heteroskedasticity does not cause bias or inconsistency in the OLS estimators, but the usual standard errors
and test statistics are no longer valid. We showed how to compute heteroskedasticity-robust standard errors
and t statistics, something that is routinely done by many regression packages. Most regression packages
also compute a heteroskedasticity-robust F-type statistic.
We discussed two common ways to test for heteroskedasticity: the Breusch-Pagan test and a special
case of the White test. Both of these statistics involve regressing the squared OLS residuals on either the
independent variables (BP) or the fitted and squared fitted values (White). A simple F test is asymptotically
valid; there are also Lagrange multiplier versions of the tests.
58860_ch08_hr_262-293.indd 286
10/18/18 4:29 PM
CHAPTER 8 Heteroskedasticity
287
OLS is no longer the best linear unbiased estimator in the presence of heteroskedasticity. When the
form of heteroskedasticity is known, GLS estimation can be used. This leads to weighted least squares as a
means of obtaining the BLUE estimator. The test statistics from the WLS estimation are either exactly valid
when the error term is normally distributed or asymptotically valid under nonnormality. This assumes, of
course, that we have the proper model of heteroskedasticity.
More commonly, we must estimate a model for the heteroskedasticity before applying WLS. The
resulting feasible GLS estimator is no longer unbiased, but it is consistent and asymptotically efficient. The
usual statistics from the WLS regression are asymptotically valid. We discussed a method to ensure that
the estimated variances are strictly positive for all observations, something needed to apply WLS.
As we discussed in Chapter 7, the linear probability model for a binary dependent variable necessarily
has a heteroskedastic error term. A simple way to deal with this problem is to compute heteroskedasticity-robust statistics. Alternatively, if all the fitted values (that is, the estimated probabilities) are strictly
between zero and one, weighted least squares can be used to obtain asymptotically efficient estimators.
Key Terms
Breusch-Pagan Test for
Heteroskedasticity (BP Test)
Feasible GLS (FGLS) Estimator
Generalized Least Squares (GLS)
Estimators
Heteroskedasticity of Unknown
Form
Heteroskedasticity-Robust F
Statistic
Heteroskedasticity-Robust LM
Statistic
Heteroskedasticity-Robust
Standard Error
Heteroskedasticity-Robust
t Statistic
Weighted Least Squares (WLS)
Estimators
White Test for Heteroskedasticity
Problems
1 Which of the following are consequences of heteroskedasticity?
(i) The OLS estimators, b^ j, are inconsistent.
(ii) The usual F statistic no longer has an F distribution.
(iii) The OLS estimators are no longer BLUE.
2 Consider a linear model to explain monthly beer consumption:
beer 5 b0 1 b1inc 1 b2 price 1 b3educ 1 b4 female 1 u
E 1 u 0 inc, price, educ, female 2 5 0
Var 1 u 0 inc, price, educ, female 2 5 s2inc2.
Write the transformed equation that has a homoskedastic error term.
3 True or False? WLS is preferred to OLS when an important variable has been omitted from the model.
4 Using the data in GPA3, the following equation was estimated for the fall and second semester s­ tudents:
trmgpa 5 22.12 1 .900 crsgpa 1 .193 cumgpa 1 .0014 tothrs
1 .55 2 1 .175 2
3 .55 4 3 .166 4
1 .064 2
3 .074 4
1 .0012 2
3 .0012 4
1 .0018 sat 2 .0039 hsperc 1 .351 female 2 .157 season
1 .0002 2
3 .0002 4
2
1 .0018 2
3 .0019 4
n 5 269, R 5 .465.
58860_ch08_hr_262-293.indd 287
1 .085 2
3 .079 4
1 .098 2
3 .080 4
10/18/18 4:29 PM
288
PART 1 Regression Analysis with Cross-Sectional Data
Here, trmgpa is term GPA, crsgpa is a weighted average of overall GPA in courses taken, cumgpa is
GPA prior to the current semester, tothrs is total credit hours prior to the semester, sat is SAT score,
hsperc is graduating percentile in high school class, female is a gender dummy, and season is a dummy
variable equal to unity if the student’s sport is in season during the fall. The usual and heteroskedasticity-robust standard errors are reported in parentheses and brackets, respectively.
(i) Do the variables crsgpa, cumgpa, and tothrs have the expected estimated effects? Which of
these variables are statistically significant at the 5% level? Does it matter which standard errors
are used?
(ii) Why does the hypothesis H0: bcrsgpa 5 1 make sense? Test this hypothesis against the two-sided
alternative at the 5% level, using both standard errors. Describe your conclusions.
(iii) Test whether there is an in-season effect on term GPA, using both standard errors. Does the
­significance level at which the null can be rejected depend on the standard error used?
5 The variable smokes is a binary variable equal to one if a person smokes, and zero otherwise. Using the
data in SMOKE, we estimate a linear probability model for smokes:
smokes 5 .656 2 .069 log 1 cigpric 2 1 .012 log 1 income 2 2 .029 educ
1 .855 2 1 .204 2
3 .856 4 3 .207 4
2
1 .026 2
3 .026 4
1 .006 2
3 .006 4
1 .020 age 2 .00026 age 2 .101 restaurn 2 .026 white
1 .006 2
3 .005 4
2
1 .00006 2
3 .00006 4
n 5 807, R 5 .062.
1 .039 2
3 .038 4
1 .052 2
3 .050 4
The variable white equals one if the respondent is white, and zero otherwise; the other independent variables are defined in Example 8.7. Both the usual and heteroskedasticity-robust standard errors are reported.
(i) Are there any important differences between the two sets of standard errors?
(ii) Holding other factors fixed, if education increases by four years, what happens to the estimated
probability of smoking?
(iii) At what point does another year of age reduce the probability of smoking?
(iv) Interpret the coefficient on the binary variable restaurn (a dummy variable equal to one if the
person lives in a state with restaurant smoking restrictions).
(v) Person number 206 in the data set has the following characteristics: cigpric 5 67.44,
income 5 6,500, educ 5 16, age 5 77, restaurn 5 0, white 5 0, and smokes 5 0. Compute
the predicted probability of smoking for this person and comment on the result.
6 There are different ways to combine features of the Breusch-Pagan and White tests for heteroskedasticity. One possibility not covered in the text is to run the regression
u^ 2i on xi1, xi2, . . . , xik, y^ 2i , i 5 1, . . . , n,
where the u^ i are the OLS residuals and the y^ i are the OLS fitted values. Then, we would test joint
significance of xi1, xi2, . . . , xik and y^ 2i . (Of course, we always include an intercept in this regression.)
(i) What are the df associated with the proposed F test for heteroskedasticity?
(ii) Explain why the R-squared from the regression above will always be at least as large as the
R-squareds for the BP regression and the special case of the White test.
(iii) Does part (ii) imply that the new test always delivers a smaller p-value than either the BP or
special case of the White statistic? Explain.
(iv) Suppose someone suggests also adding y^ i to the newly proposed test. What do you think of this idea?
58860_ch08_hr_262-293.indd 288
10/18/18 4:29 PM
CHAPTER 8 Heteroskedasticity
289
7 Consider a model at the employee level,
yi,e 5 b0 1 b1xi,e,1 1 b2xi,e,2 1 p 1 bk xi,e,k 1 fi 1 vi,e,
where the unobserved variable fi is a “firm effect” to each employee at a given firm i. The error
term vi,e is specific to employee e at firm i. The composite error is ui,e 5 fi 1 vi,e, such as in equation (8.28).
(i) Assume that Var 1 fi 2 5 s2f , Var 1 vi,e 2 5 s2v , and fi and vi,e are uncorrelated. Show that
Var 1 ui,e 2 5 s2f 1 s2v ; call this s2.
(ii) Now suppose that for e 2 g, vi,e and vi,g are uncorrelated. Show that Cov 1 ui,e, ui,g 2 5 s2f
mi
(iii) Let ui 5 m21
i g e^ 51 ui,e be the average of the composite errors within a firm. Show that
2
Var 1 ui 2 5 sf 1 s2v /mi.
(iv) Discuss the relevance of part (iii) for WLS estimation using data averaged at the firm level,
where the weight used for observation i is the usual firm size.
8 The following equations were estimated using the data in ECONMATH. The first equation is for men
and the second is for women. The third and fourth equations combine men and women.
score 5 20.52 1 13.60 colgpa 1 0.670 act
1 3.72 2
2
1 0.94 2
1 0.150 2
n 5 406. R 5 .4025, SSR 5 38,781.38.
score 5 13.79 1 11.89 colgpa 1 1.03 act
1 4.11 2
2
1 1.09 2
1 0.18 2
n 5 408, R 5 .3666, SSR 5 48,029.82.
score 5 15.60 1 3.17 male 1 12.82 colgpa 1 0.838 act
1 2.80 2 1 0.73 2
2
1 0.72 2
n 5 814, R 5 .3946, SSR 5 87,128.96.
1 0.116 2
score 5 13.79 1 6.73 male 1 11.89 colgpa 1 1.03 act 1 1.72 male ? colgpa 2 0.364 male ? act
1 3.91 2 1 5.55 2
2
1 1.04 2
n 5 814, R 5 .3968, SSR 5 86,811.20.
1 0.17 2
1 1.44 2
1 0.232 2
(i)
Compute the usual Chow statistic for testing the null hypothesis that the regression equations
are the same for men and women. Find the p-value of the test.
(ii) Compute the usual Chow statistic for testing the null hypothesis that the slope coefficients are
the same for men and women, and report the p-value.
(iii) Do you have enough information to compute heteroskedasticity-robust versions of the tests in
(ii) and (iii)? Explain.
9 Consider the potential outcomes framework, where w is a binary treatment indicator and the potential outcomes are y(0) and y(1). Assume that w is randomly assigned, so that w is independent of
[y(0),y(1)]. Let m0 5 E[y(0)], m1 5 E[y(1)], s20 5 Var[y(0)], and s21 5 Var[y(1)].
(i) Define the observed outcome as y 5 (1 2 w)y(0) 1 wy(1). Letting t 5 m1 2 m0 be the average
treatment effect, show you can write
y 5 m0 1 tw 1 (1 2 w)v(0) 1 wv(1),
where v(0) 5 y(0) 2 m0 and v(1) 5 y(1) 2 m1.
58860_ch08_hr_262-293.indd 289
10/18/18 4:29 PM
290
PART 1 Regression Analysis with Cross-Sectional Data
(ii)
Let u 5 (1 2 w)v(0) 1 wv(1) be the error term in
y 5 m0 1 tw 1 u
Show that
E(u 0 w) 5 0
What statistical properties does this finding imply about the OLS estimator of t from the simply
regression yi on wi for a random sample of size n? What happens as n S `?
(iii) Show that
(iv)
(v)
Var(u 0 w) 5 E(u2 0 w) 5 (1 2 w)s20 1 ws21.
Is there generally heteroskedasticity in the error variance?
If you think s21 2 s20, and t^ is the OLS estimator, how would you obtain a valid standard error
for t^ ?
After obtaining the OLS residuals, u^ i, i 5 1, . . . , n, propose a regression that allows consistent
estimation of s20 and s21. [Hint: You should first square the residuals.]
Computer Exercises
C1 Consider the following model to explain sleeping behavior:
sleep 5 b0 1 b1totwrk 1 b2 educ 1 b3 age 1 b4 age2 1 b5 yngkid 1 b6 male 1 u.
(i)
Write down a model that allows the variance of u to differ between men and women. The variance should not depend on other factors.
(ii) Use the data in SLEEP75 to estimate the parameters of the model for heteroskedasticity. (You
have to first estimate the sleep equation by OLS to obtain the OLS residuals.) Is the estimated
variance of u higher for men or for women?
(iii) Is the variance of u statistically different for men and for women?
C2 (i)Use the data in HPRICE1 to obtain the heteroskedasticity-robust standard errors for equation
(8.17). Discuss any important differences with the usual standard errors.
(ii) Repeat part (i) for equation (8.18).
(iii) What does this example suggest about heteroskedasticity and the transformation used for the
dependent variable?
C3 Apply the full White test for heteroskedasticity [see equation (8.19)] to equation (8.18). Using the chisquare form of the statistic, obtain the p-value. What do you conclude?
C4 Use VOTE1 for this exercise.
(i) Estimate a model with voteA as the dependent variable and prtystrA, democA, log(expendA),
and log(expendB) as independent variables. Obtain the OLS residuals, u^ i, and regress these on
all of the independent variables. Explain why you obtain R2 5 0.
(ii) Now, compute the Breusch-Pagan test for heteroskedasticity. Use the F statistic version and
report the p-value.
(iii) Compute the special case of the White test for heteroskedasticity, again using the F statistic
form. How strong is the evidence for heteroskedasticity now?
C5 Use the data in PNTSPRD for this exercise.
(i) The variable sprdcvr is a binary variable equal to one if the Las Vegas point spread for a college
basketball game was covered. The expected value of sprdcvr, say m, is the probability that the
spread is covered in a randomly selected game. Test H0: m 5 .5 against H1: m 2 .5 at the 10%
58860_ch08_hr_262-293.indd 290
10/18/18 4:29 PM
CHAPTER 8 Heteroskedasticity
291
significance level and discuss your findings. (Hint: This is easily done using a t test by regressing sprdcvr on an intercept only.)
(ii) How many games in the sample of 553 were played on a neutral court?
(iii) Estimate the linear probability model
sprdcvr 5 b0 1 b1 favhome 1 b2 neutral 1 b3 fav25 1 b4 und25 1 u
and report the results in the usual form. (Report the usual OLS standard errors and the heteroskedasticity-robust standard errors.) Which variable is most significant, both practically and statistically?
(iv) Explain why, under the null hypothesis H0: b1 5 b2 5 b3 5 b4 5 0, there is no heteroskedasticity in the model.
(v) Use the usual F statistic to test the hypothesis in part (iv). What do you conclude?
(vi) Given the previous analysis, would you say that it is possible to systematically predict whether
the Las Vegas spread will be covered using information available prior to the game?
C6 In Example 7.12, we estimated a linear probability model for whether a young man was arrested during 1986:
arr86 5 b0 1 b1 pcnv 1 b2 avgsen 1 b3 tottime 1 b4 ptime86 1 b5 qemp86 1 u.
(i)
Using the data in CRIME1, estimate this model by OLS and verify that all fitted values are
strictly between zero and one. What are the smallest and largest fitted values?
(ii) Estimate the equation by weighted least squares, as discussed in Section 8-5.
(iii) Use the WLS estimates to determine whether avgsen and tottime are jointly significant at the
5% level.
C7 Use the data in LOANAPP for this exercise.
(i) Estimate the equation in part (iii) of Computer Exercise C8 in Chapter 7, computing the
­heteroskedasticity-robust standard errors. Compare the 95% confidence interval on bwhite with
the nonrobust confidence interval.
(ii) Obtain the fitted values from the regression in part (i). Are any of them less than zero? Are any
of them greater than one? What does this mean about applying weighted least squares?
C8 Use the data set GPA1 for this exercise.
(i) Use OLS to estimate a model relating colGPA to hsGPA, ACT, skipped, and PC. Obtain the
OLS residuals.
(ii) Compute the special case of the White test for heteroskedasticity. In the regression of u^ 2i on
colGPAi, colGPA2i , obtain the fitted values, say h^ i.
(iii) Verify that the fitted values from part (ii) are all strictly positive. Then, obtain the weighted
least squares estimates using weights 1/h^ i. Compare the weighted least squares estimates for
the effect of skipping lectures and the effect of PC ownership with the corresponding OLS estimates. What about their statistical significance?
(iv) In the WLS estimation from part (iii), obtain heteroskedasticity-robust standard errors. In other
words, allow for the fact that the variance function estimated in part (ii) might be misspecified.
(See Question 8.4.) Do the standard errors change much from part (iii)?
C9 In Example 8.7, we computed the OLS and a set of WLS estimates in a cigarette demand equation.
(i) Obtain the OLS estimates in equation (8.35).
(ii) Obtain the h^ i used in the WLS estimation of equation (8.36) and reproduce equation (8.36). From
this equation, obtain the unweighted residuals and fitted values; call these u^ i and y^ i, respectively.
(For example, in Stata®, the unweighted residuals and fitted values are given by default.)
(iii) Let ŭi 5 u^ i/ "h^ i and y̆i 5 y^ i/ "h^ i be the weighted quantities. Carry out the special case of the
White test for heteroskedasticity by regressing ŭ2i on y̆i, y̆2i , being sure to include an intercept, as
always. Do you find heteroskedasticity in the weighted residuals?
58860_ch08_hr_262-293.indd 291
10/18/18 4:29 PM
292
PART 1 Regression Analysis with Cross-Sectional Data
(iv)
(v)
What does the finding from part (iii) imply about the proposed form of heteroskedasticity used
in obtaining (8.36)?
Obtain valid standard errors for the WLS estimates that allow the variance function to be
misspecified.
C10 Use the data set 401KSUBS for this exercise.
(i) Using OLS, estimate a linear probability model for e401k, using as explanatory variables inc,
inc2, age, age2, and male. Obtain both the usual OLS standard errors and the heteroskedasticityrobust versions. Are there any important differences?
(ii) In the special case of the White test for heteroskedasticity, where we regress the squared OLS
residuals on a quadratic in the OLS fitted values, u^ 2i on y^ i, y^ 2i , i 5 1, . . . , n, argue that the probability limit of the coefficient on y^ i should be one, the probability limit of the coefficient on y^ 2i
should be 21, and the probability limit of the intercept should be zero. {Hint: Remember that
Var 1 y 0 x1, . . . , xk 2 5 p 1 x 2 3 1 2 p 1 x 2 4 , where p 1 x 2 5 b0 1 b1x1 1 p 1 bkxk.}
(iii) For the model estimated from part (i), obtain the White test and see if the coefficient estimates
roughly correspond to the theoretical values described in part (ii).
(iv) After verifying that the fitted values from part (i) are all between zero and one, obtain the
weighted least squares estimates of the linear probability model. Do they differ in important
ways from the OLS estimates?
C11 Use the data in 401KSUBS for this question, restricting the sample to fsize 5 1.
(i) To the model estimated in Table 8.1, add the interaction term, e401k · inc. Estimate the equation
by OLS and obtain the usual and robust standard errors. What do you conclude about the statistical significance of the interaction term?
(ii) Now estimate the more general model by WLS using the same weights, 1/inci, as in Table 8.1.
Compute the usual and robust standard error for the WLS estimator. Is the interaction term statistically significant using the robust standard error?
(iii) Discuss the WLS coefficient on e401k in the more general model. Is it of much interest by
itself? Explain.
(iv) Reestimate the model by WLS but use the interaction term e401k · (inc − 30); the average
income in the sample is about 29.44. Now interpret the coefficient on e401k.
C12 Use the data in MEAP00 to answer this question.
(i) Estimate the model
math4 5 b0 1 b1lunch 1 b2log 1 enroll 2 1 b3log 1 exppp 2 1 u
by OLS and obtain the usual standard errors and the fully robust standard errors. How do they
generally compare?
(ii) Apply the special case of the White test for heteroskedasticity. What is the value of the F test?
What do you conclude?
(iii) Obtain g^ i as the fitted values from the regression log 1 u^ 2i 2 on math4i, math42i , where math4i are
the OLS fitted values and the u^ i are the OLS residuals. Let h^ i 5 exp 1 g^ i 2 . Use the h^ i to obtain
WLS estimates. Are there big differences with the OLS coefficients?
(iv) Obtain the standard errors for WLS that allow misspecification of the variance function. Do
these differ much from the usual WLS standard errors?
(v) For estimating the effect of spending on math4, does OLS or WLS appear to be more precise?
C13 Use the data in FERTIL2 to answer this question.
(i) Estimate the model
children 5 b0 1 b1age 1 b2 age2 1 b3 educ 1 b4 electric 1 b5 urban 1 u
and report the usual and heteroskedasticity-robust standard errors. Are the robust standard errors
always bigger than the nonrobust ones?
58860_ch08_hr_262-293.indd 292
10/18/18 4:29 PM
CHAPTER 8 Heteroskedasticity
293
(ii)
Add the three religious dummy variables and test whether they are jointly significant. What are
the p-values for the nonrobust and robust tests?
(iii) From the regression in part (ii), obtain the fitted values y^ and the residuals, u^ . Regress u^ 2 on y^ , y^ 2
and test the joint significance of the two regressors. Conclude that heteroskedasticity is present
in the equation for children.
(iv) Would you say the heteroskedasticity you found in part (iii) is practically important?
C14 Use the data in BEAUTY for this question.
(i) Using the data pooled for men and women, estimate the equation
lwage 5 b0 1 b1belavg 1 b2abvavg 1 b3female 1 b4educ 1 b5exper 1 b5exper2 1 u,
and report the results using heteroskedasticity-robust standard errors below coefficients. Are any
of the coefficients surprising in either their signs or magnitudes? Is the coefficient on female
practically large and statistically significant?
(ii) Add interactions of female with all other explanatory variables in the equation from part (i) (five
interactions in all). Compute the usual F test of joint significance of the five interactions and a
heteroskedasticity-robust version. Does using the heteroskedasticity-robust version change the
outcome in any important way?
(iii) In the full model with interactions, determine whether those involving the looks variables—
female • belavg and female • abvavg—are jointly significant. Are their coefficients practically
small?
58860_ch08_hr_262-293.indd 293
10/18/18 4:29 PM
chapter
9
More on Specification
and Data Issues
I
n Chapter 8, we dealt with one failure of the Gauss-Markov assumptions. While heteroskedasticity
in the errors can be viewed as a problem with a model, it is a relatively minor one. The presence of
heteroskedasticity does not cause bias or inconsistency in the OLS estimators. Also, it is fairly easy
to adjust confidence intervals and t and F statistics to obtain valid inference after OLS estimation, or
even to get more efficient estimators by using weighted least squares.
In this chapter, we return to the much more serious problem of correlation between the error, u,
and one or more of the explanatory variables. Remember from Chapter 3 that if u is, for whatever
reason, correlated with the explanatory variable xj, then we say that xj is an endogenous explanatory
variable. We also provide a more detailed discussion on three reasons why an explanatory variable
can be endogenous; in some cases, we discuss possible remedies.
We have already seen in Chapters 3 and 5 that omitting a key variable can cause correlation
between the error and some of the explanatory variables, which generally leads to bias and inconsistency in all of the OLS estimators. In the special case that the omitted variable is a function of an
explanatory variable in the model, the model suffers from functional form misspecification.
We begin in the first section by discussing the consequences of functional form misspecification
and how to test for it. In Section 9-2, we show how the use of proxy variables can solve, or at least
mitigate, omitted variables bias. In Section 9-3, we derive and explain the bias in OLS that can arise
under certain forms of measurement error. Additional data problems are discussed in Section 9-4.
All of the procedures in this chapter are based on OLS estimation. As we will see, certain problems that cause correlation between the error and some explanatory variables cannot be solved by using
OLS on a single cross section. We postpone a treatment of alternative estimation methods until Part 3.
294
58860_ch09_hr_294-332.indd 294
10/19/18 8:34 PM
CHAPTER 9 More on Specification and Data Issues
295
9-1 Functional Form Misspecification
A multiple regression model suffers from functional form misspecification when it does not properly
account for the relationship between the dependent and the observed explanatory variables. For example, if hourly wage is determined by log 1 wage 2 5 b0 1 b1educ 1 b2exper 1 b3exper2 1 u, but we
omit the squared experience term, exper2, then we are committing a functional form misspecification. We already know from Chapter 3 that this generally leads to biased estimators of b0, b1, and b2.
(We do not estimate b3 because exper2 is excluded from the model.) Thus, misspecifying how exper
affects log 1 wage 2 generally results in a biased estimator of the return to education, b1. The amount of
this bias depends on the size of b3 and the correlation among educ, exper, and exper2.
Things are worse for estimating the return to experience: even if we could get an unbiased estimator of b2, we would not be able to estimate the return to experience because it equals b2 1 2b3exper
(in decimal form). Just using the biased estimator of b2 can be misleading, especially at extreme values of exper.
As another example, suppose the log 1 wage 2 equation is
log 1 wage 2 5 b0 1 b1educ 1 b2exper 1 b3exper2
1 b4 female 1 b5 female # educ 1 u,
[9.1]
where female is a binary variable. If we omit the interaction term, female # educ, then we are misspecifying the functional form. In general, we will not get unbiased estimators of any of the other parameters, and because the return to education depends on gender, it is not clear what return we would be
estimating by omitting the interaction term.
Omitting functions of independent variables is not the only way that a model can suffer from
misspecified functional form. For example, if (9.1) is the true model satisfying the first four GaussMarkov assumptions, but we use wage rather than log 1 wage 2 as the dependent variable, then we will
not obtain unbiased or consistent estimators of the partial effects. The tests that follow have some
ability to detect this kind of functional form problem, but there are better tests that we will mention in
the subsection on testing against nonnested alternatives.
Misspecifying the functional form of a model can certainly have serious consequences.
Nevertheless, in one important respect, the problem is minor: by definition, we have data on all the
necessary variables for obtaining a functional relationship that fits the data well. This can be contrasted with the problem addressed in the next section, where a key variable is omitted on which we
cannot collect data.
We already have a very powerful tool for detecting misspecified functional form: the F test for
joint exclusion restrictions. It often makes sense to add quadratic terms of any significant variables
to a model and to perform a joint test of significance. If the additional quadratics are significant, they
can be added to the model (at the cost of complicating the interpretation of the model). However,
significant quadratic terms can be symptomatic of other functional form problems, such as using the
level of a variable when the logarithm is more appropriate, or vice versa. It can be difficult to pinpoint
the precise reason that a functional form is misspecified. Fortunately, in many cases, using logarithms
of certain variables and adding quadratics are sufficient for detecting many important nonlinear relationships in economics.
Example 9.1
Economic Model of Crime
Table 9.1 contains OLS estimates of the economic model of crime (see Example 8.3). We first estimate the model without any quadratic terms; those results are in column (1).
In column (2), the squares of pcnv, ptime86, and inc86 are added; we chose to include the squares
of these variables because each level term is significant in column (1). The variable qemp86 is a discrete variable taking on only five values, so we do not include its square in column (2).
58860_ch09_hr_294-332.indd 295
10/19/18 8:34 PM
296
PART 1 Regression Analysis with Cross-Sectional Data
Table 9.1 Dependent Variable: narr86
Independent Variables
(1)
(2)
pcnv
2.133
(.040)
.553
(.154)
pcnv 2
—
2.730
(.156)
avgsen
2.011
(.012)
2.017
(.012)
tottime
.012
(.009)
.012
(.009)
ptime86
2.041
(.009)
.287
(.004)
ptime86 2
—
2.0296
(.0039)
qemp86
2.051
(.014)
2.014
(.017)
inc86
2.0015
(.0003)
2.0034
(.0008)
inc86 2
—
black
.327
(.045)
.292
(.045)
hispan
.194
(.040)
.164
(.039)
intercept
.569
(.036)
.505
(.037)
Observations
R-squared
2,725
.0723
2,725
.1035
2.000007
(.000003)
Each of the squared terms is significant, and together they are jointly very significant (F 5 31.37,
with df 5 3 and 2,713; the p-value is essentially zero). Thus, it appears that the initial model overlooked some potentially important nonlinearities.
The presence of the quadratics makes interpretG o i n g F u rt h e r 9 . 1
ing the model somewhat difficult. For example, pcnv
no longer has a strict deterrent effect: the relationWhy do we not include the squares of black
and hispan in column (2) of Table 9.1?
ship between narr86 and pcnv is positive up until
Would it make sense to add interacpcnv 5 .365, and then the relationship is negative.
tions of black and hispan with some of the
We might conclude that there is little or no deterrent
other variables reported in the table?
effect at lower values of pcnv; the effect only kicks
in at higher prior conviction rates. We would have to
use more sophisticated functional forms than the quadratic to verify this conclusion. It may be that
pcnv is not entirely exogenous. For example, men who have not been convicted in the past (so that
pcnv 5 0) are perhaps casual criminals, and so they are less likely to be arrested in 1986. This could
be biasing the estimates.
Similarly, the relationship between narr86 and ptime86 is positive up until ptime86 5 4.85
(almost five months in prison), and then the relationship is negative. The vast majority of men in the
sample spent no time in prison in 1986, so again we must be careful in interpreting the results.
58860_ch09_hr_294-332.indd 296
10/19/18 8:34 PM
CHAPTER 9 More on Specification and Data Issues
297
Legal income has a negative effect on narr86 until inc86 5 242.85; because income is measured
in hundreds of dollars, this means an annual income of $24,285. Only 46 of the men in the sample
have incomes above this level. Thus, we can conclude that narr86 and inc86 are negatively related
with a diminishing effect.
Example 9.1 is a tricky functional form problem due to the nature of the dependent variable.
Other models are theoretically better suited for handling dependent variables taking on a small number of integer values. We will briefly cover these models in Chapter 17.
9-1a RESET as a General Test for Functional
Form Misspecification
Some tests have been proposed to detect general functional form misspecification. Ramsey’s (1969)
regression specification error test (RESET) has proven to be useful in this regard.
The idea behind RESET is fairly simple. If the original model
y 5 b0 1 b1x1 1 p 1 bkxk 1 u
[9.2]
satisfies MLR.4, then no nonlinear functions of the independent variables should be significant when
added to equation (9.2). In Example 9.1, we added quadratics in the significant explanatory variables. Although this often detects functional form problems, it has the drawback of using up many
degrees of freedom if there are many explanatory variables in the original model (much as the straight
form of the White test for heteroskedasticity consumes degrees of freedom). Further, certain kinds of
neglected nonlinearities will not be picked up by adding quadratic terms. RESET adds polynomials
in the OLS fitted values to equation (9.2) to detect general kinds of functional form misspecification.
To implement RESET, we must decide how many functions of the fitted values to include in an
expanded regression. There is no right answer to this question, but the squared and cubed terms have
proven to be useful in most applications.
Let y^ denote the OLS fitted values from estimating (9.2). Consider the expanded equation
y 5 b0 1 b1x1 1 p 1 bkxk 1 d1y^ 2 1 d2y^ 3 1 error.
[9.3]
This equation seems a little odd, because functions of the fitted values from the initial estimation now
appear as explanatory variables. In fact, we will not be interested in the estimated parameters from
(9.3); we only use this equation to test whether (9.2) has missed important nonlinearities. The thing to
remember is that y^ 2 and y^ 3 are just nonlinear functions of the xj.
The null hypothesis is that (9.2) is correctly specified. Thus, RESET is the F statistic for testing H0: d1 5 0, d2 5 0 in the expanded model (9.3). A significant F statistic suggests some sort of
functional form problem. The distribution of the F statistic is approximately F2,n2k23 in large samples
under the null hypothesis (and the Gauss-Markov assumptions). The df in the expanded equation (9.3)
is n 2 k 2 1 2 2 5 n 2 k 2 3. An LM version is also available (and the chi-square distribution will
have two df ). Further, the test can be made robust to heteroskedasticity using the methods discussed
in Section 8-2.
Example 9.2
Housing Price Equation
We estimate two models for housing prices. The first one has all variables in level form:
price 5 b0 1 b1lotsize 1 b2sqrft 1 b3bdrms 1 u.
[9.4]
The second one uses the logarithms of all variables except bdrms:
log 1 price 2 5 b0 1 b1llotsize 1 b2lsqrft 1 b3bdrms 1 u.
58860_ch09_hr_294-332.indd 297
[9.5]
10/19/18 8:34 PM
298
PART 1 Regression Analysis with Cross-Sectional Data
Using n 5 88 houses in HPRICE1, the RESET statistic for equation (9.4) turns out to be 4.67; this
is the value of an F2,82 random variable 1 n 5 88, k 5 3 2 , and the associated p-value is .012. This is
evidence of functional form misspecification in (9.4).
The RESET statistic in (9.5) is 2.56, with p-value 5 .084. Thus, we do not reject (9.5) at the 5%
significance level (although we would at the 10% level). On the basis of RESET, the log-log model in
(9.5) is preferred.
In the previous example, we tried two models for explaining housing prices. One was rejected by
RESET, while the other was not (at least at the 5% level). Often, things are not so simple. A drawback
with RESET is that it provides no real direction on how to proceed if the model is rejected. Rejecting
(9.4) by using RESET does not immediately suggest that (9.5) is the next step. Equation (9.5) was
estimated because constant elasticity models are easy to interpret and can have nice statistical properties. In this example, it so happens that it passes the functional form test as well.
Some have argued that RESET is a very general test for model misspecification, including unobserved omitted variables and heteroskedasticity. Unfortunately, such use of RESET is largely misguided. It can be shown that RESET has no power for detecting omitted variables whenever they have
expectations that are linear in the included independent variables in the model [see Wooldridge (200
Download